413 capabilities, each graded. 176 validated against independent external oracles. 233 honestly labelled Modelled. 4 partner-owned. Nothing is rounded up.
Verification matrix v0.41.0From the repository
176Validated
233Modelled
4Partner-owned
413capabilities
Hover or focus a cell to read its capability and the oracle that grades it.
Validated: an independent external oracle agreesModelledPartner-owned
Engine v0.41.0 · web/data/verification-matrix.json, generated from src/verification.rs · 413 rows · no scenario or seed: the engine's own ledgerRead the ledger
Free coreThe ledger, the results, the reproducibility stamp and the tests behind them are all in the open engine.
ProPackages the evidence into an audit-grade, reproducible evidence pack for a review board. See below
Custom studyAnswers your own question with the same labels, not a certification. See below
01Validated results · every compared value
Checked against someone else's answer.
Each tile draws every value the engine compared with an independent reference: published vectors, a separate flight-dynamics library, precise ephemerides, a statistics library, a standards publication, and a measured clock record held out from fitting. Hover or focus a mark for its value.
Orbits · analyticValidated
666/666
SGP4 (Simplified General Perturbations 4) orbit propagation
Every AIAA (American Institute of Aeronautics and Astronautics) reference vector reproduced, deep-space SDP4 (Simplified Deep-space Perturbations 4) cases included. Worst case 4.12 mm.
One square per compared vector, coloured by its position error.
Engine v0.41.0, build dd7509f · tests/numerical_cowell_propagator_reference.rs · deterministic, no seedNo Kshana Studio view: this comparison runs as an engine test.
Orbits · millimetres to metresValidated
0.61m Galileo/0.10m Swarm-A
Every orbit residual on one scale
From the propagator checks to fits against ESA (European Space Agency) precise ephemerides. The lunar orbiter, above the bar, is shown too.
Each residual on a logarithmic scale from a millimetre to twenty metres.
oracle · ESA precise ephemerides, the lunar orbiter's reconstructed trajectory, Orekit 12.2, AIAA vectorsHow it was checked
Engine v0.41.0 · docs/AGENCY-ORBIT-VALIDATION.md residual tables, with the two propagator checks recomputed at build dd7509f · deterministic, no seedNo Kshana Studio view: precise orbit determination runs as engine tests.
Clocks · stabilityValidated
12of 12 published values
Allan deviation
Reproduces the reference values of NIST (National Institute of Standards and Technology) Special Publication (SP) 1065, Table 31, on its own 1000-point data set.
Four deviations against averaging time, with the published values as rings.
Validated means an independent external source agrees: published reference vectors, a separate tool, or measured data. Anything else says Modelled. Pick a label to light its cells.
Capabilities graded
176of 413 validated
Hover a cell
Counts are exact, from the engine's verification matrix. Cells follow the ledger's row order.
Three words, used strictly
Kshana outputengine run
Independent oraclepublished reference
Validated
Only with an independent external oracle: published reference vectors, a separate implementation, or measured data held out from fitting. The ledger names the oracle and the test, and the matrix invariant tests enforce the rule.
Modelled
A published model or a documented modelling choice, internally tested, with no independent external oracle yet. Said out loud, not rounded up.
Partner-owned
Spacecraft bus engineering, navigation payload and antenna hardware, quantum payload hardware, and flight-hardware product assurance. Kshana claims none of them; a partner owns them.
With Pro: a mission dossier checks every requirement against real runs and keeps these labels in its verification matrix: a requirement met on a MODELLED or PARTNER basis is flagged. See a real dossier
03Validation ledger
All 413 rows, one line each.
Every row names its requirement, its status and the module that implements it. Open a row for its oracle and tests, or pick a domain to narrow it.
By domain: bar length is rows, colour is status. Pick one to filter.
ValidatedModelledPartner-owned
ValidatedPower-law noise identification by lag-1 autocorrelationallan (lag1_noise_id,lag1_acf)
Capability
allan::lag1_noise_id: the Riley and Greenhall (2004) lag-1 autocorrelation noise identifier (decimated phase with its quadratic removed, or group-averaged frequency with its line removed; repeated first differences until delta = r1/(1+r1) < 0.25 or d = dmax), returning the integer and unrounded power-law exponent alpha (+2 white PM to -2 random-walk FM) at any averaging factor; it sets the noise-type degrees of freedom of the measured clock library's device cards
Oracle
allantools 2024.6 autocorr_noise_id (A. Wallin and contributors, LGPL-3.0-or-later, run as a separate program by the fixture generator): an independent implementation of the published Riley-Greenhall algorithm, a uniquely defined quantity under the stated conventions. Validates the identifier, not the noise content of any real clock
Tests
tests/clock_library_lag1_noise_id_allantools.rs (140 cases: ten Kasdin power-law records, b = 0 to -4, as phase and frequency data, af = 1 to 64; alpha_int and d identical, worst |d alpha| 9.6e-13 and |d rho| 4.8e-13…
osnma + kshana osnma verify: an open, pure-Rust receiver-side verifier for Galileo OSNMA (SIS ICD Issue 1.1, Receiver Guidelines Issue 1.3) from I/NAV pages or a u-blox UBX-RXM-SFRBX stream: Merkle-tree public keys (DSM-PKR), the ECDSA P-256/P-521 signature of the DSM-KROOT, TESLA key-chain verification (SHA-256 or SHA3-256), MACSEQ and tag verification (HMAC-SHA-256 or CMAC-AES) for ADKD 0, 4 and 12 with cross-authentication, chain end/revocation/renewal, public-key renewal and revocation, and the alert message; per-satellite status Authenticated/Failed/Unavailable with a reason for every failure or delay. Advisory: not type-approved navigation equipment and not a navigation integrity service; no anti-spoofing performance is claimed
Oracle
The official Galileo OSNMA test vectors published by the European GNSS Service Centre as an annex to the Receiver Guidelines (Test_vectors.zip, sha256 ef9b9afc6ef9e1c57393415cf2a1dc80c4035e7c06123907ea0bb97dd3ccf370): 18 of 18 sets, 186,382 tags authenticated, 0 failed, run offline on 2026-10-10 (evidence record in kshana-pro docs/evidence). Scope: the vectors validate the cryptographic verification (Merkle, KROOT/ECDSA, TESLA chain, MACK tags, ADKD 0/4/12) and its outcomes on the published scenarios; they use only HMAC-SHA-256, a SHA-256 chain, 128-bit keys, 40-bit tags and look-up entries 33 and 34, so CMAC-AES, SHA3-256 chains and other parameters are covered by published primitive known-answer tests and synthetic tests only. The trusted-time, data-binding, expiry and header-confirmation hardening is tested by the synthetic attack tests, not by an external oracle. Not checked: certificate revocation lists, live signals or a live receiver (the UBX word layout is checked on synthetic frames only), SHA3-256 Merkle trees, freshness without an independent time. Not an EU service
Allan/modified/Hadamard deviation + power-law noise ID with χ² CIs
Oracle
NIST SP 1065 (Riley) / Stable32 reference deviations on NBS14
Tests
tests/allan_reference.rs (NBS14 vs Stable32 to 1e-4); allan::tests
ValidatedFrequency stability on a real measured clockallan
Capability
Overlapping Allan + overlapping Hadamard deviation on a real caesium standard
Oracle
Stable32 overlapping ADEV/HDEV on the measured 5071A caesium phase series (allantools)
Tests
tests/cs5071a_reference.rs (real 5071A Cs vs H-maser, 556 990 pts, 16 averaging factors vs Stable32 to 1e-3; data-gated via scripts/fetch_cs5071a.sh)
ValidatedAllan estimator parity on the canonical Stable32 reference seriesallan
Capability
Overlapping Allan + modified Allan + time deviation across the full AF ladder
Oracle
Stable32 reference deviations for PHASE.DAT (Riley; the standard regression series)
Tests
tests/phasedat_reference.rs (Stable32 PHASE.DAT, 139 averaging factors, OADEV/MDEV/TDEV to 1e-3; data-gated via scripts/fetch_phasedat.sh)
ValidatedExtended-range frequency stability (Theo1 / TOTDEV)allan
Capability
Theo1 and total deviation (TOTVAR) — extended-range long-tau stability estimators reaching ~75% of the record where the Allan deviation gives out near ~50%; the bias-removed ThêoH hybrid (allan::theoh_curve) built on them stays MODELLED
Oracle
allantools 2024.06 — an independent third-party frequency-stability library — theo1 (NIST SP 1065 eq 30) and totdev (NIST SP 1065 eq 25) on the hermetic NIST SP 1065 §12.4 LCG data set; regenerable offline via tests/fixtures/theo1_totvar/generate_theo1_totvar_reference.py
Tests
tests/theo1_totvar_reference.rs (Theo1 + TOTDEV on the NIST SP 1065 §12.4 1000-point LCG data set, 6 + 6 averaging factors vs allantools 2024.06 to <1e-9); allan::tests (white-FM closed-form tracking; TOTVAR=ADEV…
ValidatedMaximum Time Interval Error (MTIE) — telecom wander metricallan (mtie,mtie_curve)
Capability
MTIE(τ): the maximum peak-to-peak time-error swing over any sliding window of τ = m·tau0, the ITU-T G.810/G.823/G.8261 wander statistic synchronisation-network limits (MTIE masks) are written against — an extremal (max/min) figure distinct from the RMS Allan family
Oracle
allantools 2024.06 — an independent third-party frequency-stability library — mtie() on the hermetic NIST SP 1065 §12.4 LCG phase series; MTIE is a pure max/min statistic, so the estimator output is bit-exact against allantools on the identically-built phase array (the committed 15-significant-figure reference constants sit within 1 ULP). Regenerable offline via tests/fixtures/mtie/generate_mtie_reference.py
Tests
tests/mtie_reference.rs (MTIE on the hermetic NIST SP 1065 §12.4 1000-point LCG phase series, 9 averaging factors m=1..256 vs allantools 2024.06 mtie to <1e-9, observed ≤4e-15); allan::tests (hand-derived peak-to-peak…
ValidatedModified Allan / Time deviation (MDEV / TDEV)allan (modified_adev,time_deviation)
Capability
MDEV(τ): the overlapping modified Allan deviation (second-difference sliding-window estimator that separates white- from flicker-phase noise), and TDEV(τ) = τ/√3·MDEV(τ), the ITU-T G.811/G.812/G.823 time-domain wander statistic the sync masks are written against
Oracle
allantools 2024.06 — an independent third-party frequency-stability library — mdev() and tdev() on the hermetic NIST SP 1065 §12.4 LCG phase series (same series as the Theo1/TOTDEV/MTIE rows), computing the same uniquely-defined estimators; matched to <1e-9 relative. Regenerable offline via tests/fixtures/mdev_tdev/generate_mdev_tdev_reference.py. (The data-gated phasedat_reference.rs also checks these against Stable32 to 1e-3; this is the tight, always-on hermetic cross-check)
Tests
tests/mdev_tdev_reference.rs (MDEV + TDEV on the hermetic NIST SP 1065 §12.4 1000-point LCG phase series, 8 averaging factors m=1..200 vs allantools 2024.06 mdev/tdev to <1e-9 relative, plus the τ/√3 identity on the…
ValidatedOptical-clock frequency stability on a real measured curvequantum_trade (qparams_from_adev_curve),powerlaw
Capability
Power-law (white-FM + red-noise-floor) NNLS recovery from a published measured ⁸⁸Sr optical-clock-transition Allan deviation — reproducing σ_y(τ) and the headline 4.7e-16/√τ short-τ scaling with a genuine measured long-τ floor, rather than the synthesised optical-class floor holdover.rs otherwise assumes
Oracle
Norcia, Young, Eckner, Oelker, Ye, Kaufman, Science 366:93 (2019), Fig. 4 measured ADEV; curve vendored verbatim from Zenodo 10.5281/zenodo.3382347 (CC-BY-4.0). Scoped to reproducing the published measured stability curve — the clock-class holdover-to-threshold device figures stay MODELLED
Tests
tests/optical_clock_adev_reference.rs (Norcia et al. ⁸⁸Sr tweezer-clock σ_y(τ), 8 averaging times 0.92–117.76 s vendored verbatim under CC-BY-4.0: NNLS reconstructs the curve to ~10% RMS / ≤21% worst point; recovered…
ARAIM multiple-hypothesis solution separation per the ARAIM Airborne Design Document (ADD) v4.2 (araim_reference::add_v42_protection_levels and add_v42_protection_levels_ecef: subset determination by the P_THRES and FC_THRES rules, VPL, HPL, the effective monitor threshold EMT and sigma_acc); chi-squared snapshot RAIM fault detection with single-satellite exclusion (raim::snapshot_raim, raim::snapshot_raim_fde) and slope-based HPL/VPL on the geodetic local level; and the SBAS DO-229E protection-level combination (sbas::sbas_protection_level) on the L1 service. NOT covered, stated plainly: the uniform-sigma convenience functions raim::araim_raim and raim::araim_dual_raim are NOT ADD-conformant (on matched inputs they miss Stanford MAAST by 0.75 m VPL / 4.04 m HPL and 0.39 m VPL / 1.93 m HPL, a finding); the dual-frequency multi-constellation (DFMC) L5 SBAS comparison is a finding (308 user-epoch pairs with no GEO in view are protected by Kshana and not by MAAST, whose GEO-reception gate Kshana does not model); raim::solution_separation_raim has no external comparison
Oracle
Three independent tools, pre-registered (ba7a01a7) before any fixture or oracle output existed, tolerances fixed then. (1) ARAIM: Stanford MAAST for ARAIM 2 (commit ab70e2a3, BSD-3) mhss_raim_baseline_v5.m, run as a separate program under GNU Octave 8.4 with a driver-side unique() compatibility shim (disclosed, no algorithm change), at the ADD TOL_PL of 0.05 m: the Rust reproduces the ADD reference implementation on 270 real cases. Disclosed: the Kshana ADD path was written after the pre-registration, following the ADD as MAAST implements it; an independent hold-out by the reviewer (270 new user geometries and a second integrity support message, MAAST re-run against the frozen code) passed at the same bar. (2) Snapshot RAIM: RTKLIB v2.4.2-p13 (BSD-2) estpos/valsol/raim_fde in a separate C harness for detection and exclusion; the slope PL from RTKLIB matinv/xyz2enu plus SciPy 1.18.1 chi2/ncx2, where both sides evaluate the same closed form (Brown; Parkinson and Axelrad), so the PL check covers the implementation, not the formula. (3) SBAS: Stanford MAAST (commit 7d32b049) usr_vhpl, unmodified, on MAAST own recorded WAAS broadcasts: L1 levels within 1e-4 m after the K rescaling and an equal protected set. The same comparisons produced the findings stated in the capability: the uniform-sigma ARAIM functions are not ADD-conformant, and the L5 set equality fails on 308 pairs with no GEO in view (the levels on the 3267 pairs both tools protect agree to 1e-13 m). The DO-229E/DO-316 K-factors are transcribed constants and the real geometry is an input; neither is counted as an oracle. 0.30 revision: snapshot_raim protection levels moved from the radial to the geodetic local level (up to 0.31 m HPL and 0.14 m VPL at 16 deg N)
Tests
tests/integrity_araim_stanford_oracle.rs::araim_mhss_matches_stanford_maast_add_v4_2 (270 real-geometry cases, Celestrak GPS+Galileo and IGS SP3: worst |dVPL| 2.2e-3 m, |dHPL| 1.1e-3 m, |dEMT| 2.1e-10 m against the ADD…
ValidatedNumerical Cowell propagator & force modelpropagator,forces
Capability
Cowell numerical propagator with a hierarchical force model (two-body, J2–J6 zonal, Sun/Moon third-body, cannonball SRP, exponential drag); RK4 step-doubling / DP5(4) / DOP853
Oracle
Orekit 12.2 (CS GROUP, Apache-2.0) NumericalPropagator/DormandPrince853 — an independent library, a different integrator and spherical-harmonic recursion. The conservative tiers (two-body → J2–J6 zonal → Sun/Moon third-body → cannonball SRP) agree to sub-metre over a 24 h arc, validating the integrator + force algebra; the drag tier and the absolute Sun/Moon-ephemeris / density input fidelity stay MODELLED (characterisation only)
Tests
tests/numerical_cowell_propagator_reference.rs (275 epochs = 11 cases × 25 hourly states, LEO+GTO, vs Orekit 12.2 DormandPrince853; conservative tiers T1–T5 worst |Δr| 0.08 m over 24 h; drag tier characterised at 333 m)
SRP lit fraction with the WGS-84 ellipsoid as the occulting body (apparent Earth radius from the ellipsoid limb towards the Sun, dual-disk lens formula; srp_shadow::oblate_conical_shadow), opt-in on the Cowell force model through propagator::ForceModel::srp_shadow (default: the spherical conical shadow, unchanged)
Oracle
Orekit 12.2 (Apache-2.0) SolarRadiationPressure.getLightingRatio with a WGS-84 and a spherical OneAxisEllipsoid occulting body, run as a separate program with the Sun fixed at the same inertial position. Bars pre-registered in xval/srp-oblate-shadow-orekit/PREREGISTRATION.md (c6b751e9): 1e-6 on both, and a > 1e-2 oblate-vs-sphere spread; all met. The ellipsoid axis is the inertial z axis on both sides (no precession-nutation)
Tests
tests/srp_oblate_shadow_orekit_oracle.rs (2400 geometries across the shadow edge, 884 in penumbra, LEO to GEO, Sun declination 0 and +-23.44 deg, 8 azimuths: worst |dnu| 7.41e-9 for WGS-84 and 2.24e-9 for the spherical…
ValidatedBox-wing and ECOM2 solar radiation pressure, with partials (opt-in)srp_boxwing,propagator,precise_od
Capability
Non-cannonball SRP: a box-wing flat-plate model (Vallado eq. 8-44; single-sided body panels, sun-pointing wings) with the Galileo FOC, IOV and IOV end-of-life tables of the GSC satellite metadata, in nominal GNSS yaw steering, with analytic partials for the radiation factor and every panel's specular and diffuse coefficients; and the ECOM2 empirical model (D2/B1) with its partials. The box-wing is opt-in on the Cowell propagator (propagator::ForceModel::box_wing_srp, setting its srp_boxwing field) and on the precise orbit-determination model (precise_od::PreciseForceModel::box_wing_srp), replacing the cannonball term with cr as the radiation factor; both default to None, leaving every existing path byte-identical. ECOM2 is not wired into a force model
Oracle
Orekit 12.2 (Apache-2.0) BoxAndSolarArraySpacecraft with FixedPanel and PointingPanel, SolarRadiationPressure(dRef = 1 au, pRef = 1361/c), GenericGNSS nominal yaw and ECOM2(nD = 2, nB = 1), run as a separate program with the Sun at fixed inertial positions; the radiation-factor derivative from Hipparchus Gradient, per-panel derivatives as central differences of Orekit (Orekit has none). Bars pre-registered in xval/srp-boxwing-orekit/PREREGISTRATION.md (42e864bb), before any Orekit output existed: 1e-12 relative (1e-11 end to end, 1e-9 for the differences); all met (RESULTS.md). VALIDATED covers the model evaluation only: the GSC tables are transcribed inputs, and the adequacy of a box-wing for real Galileo orbits, thermal re-emission, the modified low-beta yaw law and eclipse turns are not covered; the Galileo SRP prediction stays MODELLED
Tests
tests/srp_boxwing_orekit_oracle.rs (252 Galileo-orbit geometries, 7 Sun cases including perihelion and aphelion distances, 100 random attitudes; flat-plate sum worst 2.05e-15 relative over 1056 cases; end to end with…
SRP lit fraction as the un-occulted fraction of the limb-darkened solar flux (Eddington-approximation law I(mu)/I(1) = 2/5 + 3/5 mu, Gauss-Legendre quadrature) over a spherical or WGS-84 Earth (srp_shadow::limb_darkened_shadow, the ShadowModel variant OblateEarthLimbDarkened)
Oracle
None external: Orekit 12.2 has no limb-darkened lighting ratio. Checked against closed forms of the same law only
Tests
srp_shadow::tests (the annular closed form, the uniform law reproducing the conical shadow, full sun and umbra limits)
A port of the NRLMSISE-00 empirical thermosphere (gtd7 and gtd7d: species densities, total mass density with anomalous oxygen, temperatures; the 24 switches; daily-Ap and 3-hourly ap-array modes) from the public-domain C version of the NRL Fortran
Oracle
pymsis 0.13.0 (MIT) version=0, which runs the NRL Fortran GTD7D in single precision, and the NRLMSISE-00 C port (release 20190709, public domain) compiled as a separate program, plus the C distribution's 17 published test outputs. Bars pre-registered in xval/nrlmsise00-pymsis/PREREGISTRATION.md (a6504ec9) before the port existed; all met (RESULTS.md). Validates the port, not the model's accuracy against measured density
Tests
tests/nrlmsise00_reference.rs (720 grid points, 90 to 1000 km, 4 latitudes, 2 local times, 3 days, low/moderate/high activity: density within 1.866e-6 relative and temperature within 1.97e-4 K of pymsis; bit-identical…
forces::DensityModel (Exponential default, Nrlmsise00 { f107a, f107, ap }) and propagator::ForceModel::drag_density: quadratic drag against the co-rotating atmosphere with the NRLMSISE-00 density at the WGS-84 geodetic point (Earth rotation angle with UT1 = TT, local solar time UT + lon/15), constant space weather per model; default off and bit-identical
Oracle
Orekit 12.2 (Apache-2.0) NumericalPropagator/DormandPrince853 with DragForce over Orekit's own NRLMSISE00 implementation, run as a separate program with the same Earth-rotation-angle frame, WGS-84 shape, local-solar-time convention (a mean Sun) and TT as UT; the convention setup was checked oracle against oracle (Orekit vs pymsis, 5.2e-5 in density). Bars pre-registered in xval/drag-tier-orekit/PREREGISTRATION.md (f380f42c): 0.5 m + 1e-3 of the drag displacement, and two > 100 m sensitivity checks; all met. Not covered: the model's accuracy against measured density, time-varying space weather, a GMAT leg
Tests
tests/drag_tier_orekit_oracle.rs (4 cases x 25 hourly states over 24 h, circular 51.6 deg, 400 km at low/moderate/high activity and 300 km: worst |dr| 0.0303 m and |dv| 3.36e-5 m/s against 3.3 to 141.8 km of drag…
ValidatedDefault exponential drag against an independent implementation (equatorial)forces (atmospheric_density,drag_accel),propagator
Capability
The default drag of the Cowell propagator: quadratic drag against the co-rotating atmosphere with forces::atmospheric_density, the piecewise-exponential table of Vallado Table 8-4 at the spherical altitude (propagator::ForceModel::drag with the default Exponential variant of forces::DensityModel)
Oracle
GMAT R2022a (NASA GSFC, Apache-2.0) GmatConsole with the Exponential atmosphere (the same 28-band table), point-mass Earth and PrinceDormand78, run, never vendored. Bars pre-registered in xval/drag-exponential-gmat/PREREGISTRATION.md (d870eb71) on the drag displacement of each side (drag minus no-drag run), which cancels the two Earth mu values and integrators; all met (RESULTS.md). GMAT takes geodetic height and kshana spherical altitude: they coincide only on the equator, so VALIDATED covers equatorial orbits; off the equator the models differ by design (10.4 % of 24 h displacement at 51.6 deg, 400 km). GMAT's public build has no NRLMSISE-00 (plugin absent), so the NRLMSISE-00 tier has the Orekit oracle only. The table's accuracy against measured density is not covered
Tests
tests/drag_exponential_gmat_oracle.rs (equatorial cases E400, E300 and E250x600 at J2000.0, 25 hourly states over 24 h: the drag-induced displacement agrees with GMAT's to at most 0.016 of the bar 0.5 m + 1e-3|D|…
ModelledDE440 Sun and Moon in the Cowell force model (opt-in)propagator (ForceModel::with_ephemeris,ForceEphemeris),ephem_provider
Capability
Opt-in ephemeris tier for the third-body and SRP terms of the numerical Cowell ForceModel: the geocentric Sun and Moon from any EphemerisProvider (a DE440 SPK kernel through KernelEphemeris) at TDB = TT + (TDB - TT), refusing a provider that cannot serve the epoch and counting any fallback to the analytic series; default off
Oracle
Orekit 12.2 (Apache-2.0) NumericalPropagator/DormandPrince853 with ThirdBodyAttraction and SolarRadiationPressure on Orekit's own DE440 reader (orekit-data lnxp1990.440), run as a separate program; kshana reads a bit-for-bit, SPICE-checked cut of NAIF de440s.bsp. Bars pre-registered in xval/propagator-de440-orekit/PREREGISTRATION.md (6386b3a2). The orbit bar (0.5 m) and the sensitivity control are met, but the Sun/Moon position bar (3 m / 0.2 m) missed at 6.39 m / 0.244 m. SPICE reading the same kernel at its exact TDB misses Orekit by the same amount (6.25 m / 0.240 m), so the gap lies between NAIF's de440s path and Orekit's DE440 reader or time scale; cause not isolated. Recorded and pinned, so the row stays MODELLED
Tests
tests/propagator_de440_orekit_oracle.rs (5 cases x 25 hourly states over 24 h, LEO, GEO and HEO, third body and third body + SRP: worst |dr| 0.0170 m, |dv| 1.84e-5 m/s; analytic-series control misses by 9.96 m; Sun…
The diurnal S1 air-pressure tide as a geopotential perturbation: Biancale and Bode (2006) model N1, degrees 2 to 6, through the IERS 2010 Eq. 6.21 load formula (tides::atmospheric_tide_s1, s1_air_tide_acceleration), opt-in on the Cowell and precise force models (air_tide_s1, default off)
Oracle
None external for the geopotential effect. The 24 coefficients are transcribed from GFZ STR06/01 Table 1 model N1 and were re-checked row by row against the report (SHA-256 890865ae); degrees 0, 1, 7 and 8 are not used. Ray and Ponte (2003) print no spherical-harmonic table, so no second source was compared
Tests
tides::tests (the S1 term is diurnal, turns 90 deg in a quarter day, and has the load-factor amplitude of each row; physical magnitude at LEO); propagator::tests::the_s1_air_tide_flag_adds_exactly_the_s1_acceleration
ValidatedS2 atmospheric tide coefficients in the IERS conventiontides (s2_air_waves,atmospheric_tide)
Capability
Ray (2001) S2 air-pressure harmonics (n <= 4, prograde and retrograde) converted to the fully-normalised, sine-lead form of IERS 2010 Eqs. 6.20-6.21 (amplitude divided by N_nm, epsilon = 90 deg - psi) and applied as the S2 term of tides::atmospheric_tide
Oracle
Biancale and Bode, GFZ Scientific Technical Report STR06/01 (2006), Table 5, model A2: Ray (2001) published in fully-normalised harmonics with sine-lead phases, an independent rendering of the same model. Pre-registered in xval/s2-air-tide-fix/PREREGISTRATION.md (d2c3d297) with the fix it validates. Covers the coefficient conversion only; the air tide's geopotential effect has no external oracle
Tests
tests/s2_air_tide_gfz_table5.rs (21 of 21 waves agree with the published model A2: worst 0.00050 cm, 0.050 deg against bars of 0.0006 cm, 0.1 deg); tests/tides_iers.rs (S2 amplitude band)
ValidatedTesseral geopotential in the Cowell propagatorpropagator (Geopotential,EarthFixedFrame),gravity_sh
Capability
Opt-in EGM2008 tesseral field (to degree/order 70) as the central gravity of the numerical Cowell ForceModel, evaluated in an Earth-fixed frame rotated by the IAU 2000 Earth rotation angle (or the CIO GCRS-to-ITRS chain); default off, so no existing output moves
Oracle
Orekit 12.2 (CS GROUP, Apache-2.0) NumericalPropagator/DormandPrince853 with HolmesFeatherstoneAttractionModel over the same coefficients (tools/egm2008_to70.gfc) in a body frame rotated by the same Earth-rotation-angle formula, run as a separate program. Bars pre-registered (bdab936, amended before any comparison in 2c91443): |dr| <= 0.5 m and |dv| <= 5e-4 m/s, and both negative controls missing by > 100 m; all met, bars then tightened to 0.1 m and 1e-4 m/s. Not validated: Earth-orientation fidelity (both sides use UT1 = TT without precession-nutation in the Era frame), the Cio frame option (internal check only), degrees above 70
Tests
tests/cowell_tesseral_orekit_oracle.rs (6 cases x 25 hourly states over 24 h: EGM2008 8x8 and 70x70 on LEO 45 deg, polar 89.9 deg, sun-synchronous 98.2 deg, GTO, and 70x70 with Sun/Moon and SRP; worst |dr| 0.0355 m…
ModelledHigh-degree spherical-harmonic gravity (EGM2008 to 2190)gravity_sh,propagator (FieldEvaluator)
Capability
Underflow-safe, pole-safe evaluation of a fully-normalized field to degree 2190 (SphericalHarmonicField::acceleration_hf / potential_hf, Holmes-Featherstone 2002 modified forward column with the gradient free of 1/cos phi), a streaming ICGEM .gfc loader (from_gfc_file / from_gfc_reader; the full EGM2008 is fetched at run time, not bundled), and the Holmes-Featherstone evaluator selectable on the opt-in Cowell Geopotential
Oracle
Orekit 12.2 (Apache-2.0) HolmesFeatherstoneAttractionModel and pyshtools 4.14.1 (BSD-3-Clause) MakeGravGridPoint, run as separate programs; bars pre-registered in xval/high-degree-gravity/PREREGISTRATION.md (73b2c75d) and xval/cowell-tesseral-200-orekit/PREREGISTRATION.md (876be39d). Orekit agrees everywhere it is defined, but pyshtools misses its pre-registered bar at the two near-pole points (up to 7.3e-10 m/s^2; a 50-digit mpmath check puts the error in pyshtools), there is no oracle value exactly on the polar axis, and the 24 h orbit's pre-registered degree-70 sensitivity control missed (0.726 m against > 1 m). Each miss is recorded, not tuned away, so the row stays MODELLED
Tests
tests/gravity_sh_high_degree.rs (Orekit at 8 off-axis points: max |da| 2.2e-17 m/s^2 at N = 70, always run; 3.5e-17, 3.9e-17 and 7.3e-17 m/s^2 at N = 200, 360, 2190, data-gated on KSHANA_EGM2008_GFC)…
ValidatedDOP853 integratorintegrator,propagator
Capability
Hairer DOP853 8(5,3) explicit Runge-Kutta with the combined 5th/3rd-order error estimator, the SciPy/Hairer step controller, automatic initial step, a fail-closed step guard and 7th-order dense output; selectable as the propagator Integrator variant Dop853 / propagator::propagate_dop853
Oracle
SciPy 1.18.1 (BSD-3-Clause) solve_ivp(method=DOP853) and its single-step, error-norm, dense-output and initial-step routines, an independent implementation; exact two-body states from mpmath 1.3.0 at 40 digits. Bars pre-registered in xval/dop853-scipy/PREREGISTRATION.md (8b23409) before the first comparison; all met (xval/dop853-scipy/RESULTS.md). SciPy's first step is given to both solvers in the full runs; the initial-step routine is compared separately. Not covered: RKF7(8), stiffness detection, event location
Tests
tests/dop853_scipy_oracle.rs (8 single steps: state within 5.55e-15 relative, FSAL derivative 3.46e-15, dense output 1.99e-14, accept/reject decisions identical; 3 full runs at rtol = atol = 1e-12 (LEO 58000 s, e = 0.7…
Elliptic Kepler solver for all 0 <= e < 1 and all mean anomalies: Markley (1995) cubic starter and fifth-order correction, a cancellation-free residual (1-e)E + e(E - sin E) - M, and guarded Newton polishing (kepler::solve_kepler_markley, propagator::solve_kepler_robust)
Oracle
mpmath 1.3.0 (BSD-3-Clause) 50-digit roots from the exact binary inputs. Bars pre-registered in xval/kepler-mpmath/PREREGISTRATION.md (73bab12), stated relative to |E| with the e -> 1, M -> 0 conditioning argument; all met (xval/kepler-mpmath/RESULTS.md). The legacy solvers (solve_kepler_checked, orbit::solve_kepler, ephem::solve_kepler) are numerically unchanged, because committed outputs depend on them, and are not covered by this row
Tests
tests/kepler_mpmath_oracle.rs (405 grid points, e from 0 to 0.999999, M in [-pi, pi] down to 1e-12: worst |E - E_ref| 0.898 eps|E| against a bar of 8 eps|E|; 9 wound-M points at worst 0.013 of their bar; no Newton…
ModelledRange-rate and az/el batch orbit determination with station maskingorbit_determination (determine_orbit_batch_tracking,propagate_with_stm,is_visible),batch_ls (gauss_newton_analytic)
Capability
Recover an epoch state from mixed range, range-rate and azimuth/elevation tracking above an elevation mask by Gauss-Newton on H = (dh/dx(t)) Phi(t, t0): analytic measurement partials and a variational state-transition matrix with analytic two-body + J2 partials (RK4-integrated, not a closed-form STM)
Oracle
Orekit 12.2 (Apache-2.0) BatchLSEstimator with custom instantaneous geometric RangeRate and AzEl measurements (cross-checked against Orekit's TopocentricFrame), run as a separate program on identical masked, seeded observations. Bars pre-registered in xval/od-measurements-orekit/PREREGISTRATION.md (acfd14c). Every bar met except B4, the noiseless state agreement of 1 cm, missed on 2 of 4 scenarios (1.46, 1.74 cm): Orekit's own noiseless estimate sits that far from its simulated truth (cause on the Orekit side not isolated). Recorded and pinned, so the row stays MODELLED. Not modelled: light time, aberration, refraction, Earth rotation, a sequential filter for these measurement types
Tests
tests/od_tracking_measurements_reference.rs (8 scenarios: range-rate, az/el and mixed, LEO and MEO, noiseless and seeded sigma 5 m / 5 mm/s / 1e-4 rad, 5 and 10 deg masks, against Orekit 12.2: visible sets identical…
Map SP3 accuracy codes (header ++ exponents in 2^n mm, the %f base, per-record base^n mm standard deviations) to per-observation position sigmas for the precise orbit-determination engine (sp3::parse_sp3_accuracy, Sp3Accuracy::position_sigma_m)
Oracle
None external: the weights are a modelling choice (the SP3 format defines the code only as the producer's accuracy indicator; it is read here as a per-axis one-sigma). The existing recorded agency runs are unchanged
Tests
src/sp3.rs unit tests; tests/agency_galileo.rs::galileo_e11_sp3_accuracy_weighted_fit (committed 8 h ESOC fixture, sigma 0.032 m from header exponent 5: Tier-1 3D RMS 0.132 m unchanged, Tier-2 0.035 m against 0.070 m…
ValidatedPlanetary gravity fields from PDS SHADR tablesgravity_pds,gravity_sh
Capability
Read PDS SHADR spherical-harmonic tables (from_shadr, on SphericalHarmonicField in src/gravity_pds.rs: unit conversion, C00 = 1 supplied, completeness and normalisation checks) and evaluate the body-fixed gravity of Mars GMM-3 (d/o 120) and the GRAIL lunar field GL0660B (d/o 660)
Oracle
pyshtools 4.14.1 (BSD-3-Clause) reading the same PDS files and evaluating point gravity, run as a separate program, and the published PDS product headers and labels. Bars pre-registered in xval/planetary-gravity-pyshtools/PREREGISTRATION.md (6d2674d); all met (RESULTS.md). Data: NASA PDS Geosciences Node, public domain; degree-20 crops bundled with NOTICE, full files fetched at test time
Tests
tests/planetary_gravity_pds.rs (degree-20 crops always run: worst |dg|/|g| 9.7e-16, perturbation-relative 3.4e-11; GMM-3 d/o 120 over the whole sphere 3.7e-15 / 6.3e-12 and GL0660B d/o 660 at |lat| <= 60 deg 5.0e-14 /…
ModelledAttach a published gravity field to a bodygravity_pds,body,propagator,mars_frame
Capability
with_gravity_field (on Body, in src/gravity_pds.rs): attach any loaded spherical-harmonic field (for example GMM-3 or GL0660B) to a Body so the propagator evaluates it body-fixed; opt-in, default bodies unchanged
Oracle
Internal consistency only. Each published field is defined in its producer's body-fixed frame (GMM-3 in the Konopliv et al. 2006 Mars frame, GL0660B in the DE421 lunar principal-axis frame), while the field is evaluated in the IAU frame; the offset is not modelled
Recover an epoch state [r,v] from ground-station ranges via a Gauss–Newton batch differential corrector and a sequential filter, over a two-body+J2 force model
Oracle
Orekit 12.2 (CS GROUP, Apache-2.0) BatchLSEstimator (Levenberg–Marquardt) + KalmanEstimator (EKF) — an independent third-party estimation library on a matched force model; recovered epoch state + post-fit residual RMS agree to <1e-3 m noiseless and <3 m at a 5 m noise floor
Tests
tests/batch_sequential_orbit_determination_reference.rs (8 scenarios: 6 noiseless LEO/MEO/eccentric/3–4-station + 2 σ=5 m, vs Orekit 12.2; worst batch |Δr| 1e-3 m / |Δv| 1e-6 m/s, sequential |Δr| 0.9 m)
ANISE 0.10 (Nyx Space, MPL-2.0) converged-Newtonian aberration light time (Aberration::CN; SPICE spkapo-equivalent) over JPL DE440 — an independent Rust SPICE implementation; kshana's fixed-point retarded solver matched to sub-nanosecond. The broader Doppler / Shapiro / reduced-dynamic OD figures stay MODELLED
Tests
tests/deep_space_mars_radiometric_reference.rs (24 legs over 8 epochs 2020–2027 vs ANISE DE440; worst |Δτ| 1.03e-9 s, |Δrange| 0.31 m at up to 2.5 AU)
ValidatedBroadcast-ephemeris satellite position (multi-GNSS RINEX)rinex (parse_nav,RinexEphemeris::sv_position_ecef)
Capability
IS-GPS-200 / Galileo-OS / BeiDou-OS broadcast-ephemeris Keplerian → ECEF satellite position from a parsed RINEX-3 navigation record
Oracle
RTKLIB v2.4.2-p13 eph2pos() compiled from C source (T. Takasu, BSD-2-Clause) — an independent IS-GPS-200/SIS-ICD implementation, fed the identical RINEX-3 nav records; satellite ECEF position matched to ~62 nm. (SP3 precise-ephemeris interpolation is validated separately — see 'SP3 precise-ephemeris interpolation')
IGS-standard Lagrange interpolation of SP3 precise-ephemeris satellite positions with the per-node Earth-rotation correction (rotate each node by ω⊕·(t_node−t) into the query instant's Earth-fixed frame before the polynomial fit)
Oracle
RTKLIB peph2pos() compiled from C source (preceph.c; T. Takasu, BSD-2-Clause) — the de-facto IGS reference, an independent implementation; kshana's interpolator (now carrying the same Earth-rotation node correction) matched to ~15 nm on a real SP3 product, down from ~5.5 cm before the correction
Tests
tests/sp3_interp_reference.rs (72 off-node SV-epoch cases / 6 satellites vs RTKLIB peph2pos; worst per-axis |Δ| 1.5e-8 m)
GPS CNAV message types 10/11/30, Galileo I/NAV word types 1-5 in nominal even/odd pages and Galileo F/NAV page types 1-4: ICD field packing and quantisation, CRC-24Q, rate-1/2 K=7 convolutional coding (G2 inverted for Galileo), block interleaving and sync patterns, with matching test decoders
Oracle
Internal round trip and ICD known answers; no third-party decoder has been run (the RTKLIB/GNSS-SDR harness reading KSHANA_NAVMSG_ORACLE_JSON is data-gated). MODELLED
Tests
gnss_navmsg::tests (CRC-24Q check value 0xCDE703; convolutional impulse response; Viterbi corrects injected errors; layout widths); tests/gnss_navmsg_roundtrip.rs (field start bits equal the ICD bit numbers; encode ->…
ValidatedIONEX global ionosphere map reader and interpolation on a real CODE GIMionex (parse_ionex,parse_ionex_file,IonexFile::vtec_at,IonexFile::slant_tec_at,interpolate_tec_in_time,TecGrid::vtec_at)
Capability
IONEX 1.0 reader (header, fixed-column I5 values, TEC and RMS maps, absolute epochs, missing values) with bilinear spatial interpolation, the format's nearest / consecutive / rotated-map time interpolation (eqs. 1-3) and thin-shell slant TEC
Oracle
The CODE GIM's printed values (exact at every node; the map is fetched at test time, not vendored, its licence being unconfirmed) and an independent Python stdlib re-implementation of the IONEX 1.0 interpolation formulas committed with the fixture (bars pre-registered 79beaab, 1e-9 TECU). Third-party check (B6, pre-registered, first run): RTKLIB v2.4.2-p13 iontec (BSD-2-Clause, run as a separate tool) at 6885 query points agrees with IonexFile::vtec_at to 3.6e-13 TECU in both interpolation modes against a 1e-5 TECU bar. Data-gated: CODE GIM fetched at test time, SHA-pinned; RTKLIB reads the full file because its getindex rejects the lat crop. tests/fixtures/ionex_code_gim/fetch.sh fetches it and checks the SHA-256, the test fails on any other bytes, and run_rtklib.sh is the documented reproduce path (network access needed). Promoted to VALIDATED by the owner's decision on data-gated evidence (2026-10-13)
Tests
tests/ionex_code_gim_oracle.rs (committed synthetic IONEX map vs the formula reference: 20 075 nodes exact, interpolation max |Δ| 3.5e-13 TECU; data-gated on KSHANA_IONEX_CODE_GIM_DIR, the real CODE final GIM…
ModelledHatch carrier smoothing of code on real RINEX observationshatch (HatchFilter),pvt (assemble_epoch_precise_smoothed,CarrierSmoother)
Capability
Divergence-free Hatch filter (ionosphere-free code carried by the ionosphere-free carrier, window N epochs, restart on loss of lock, geometry-free jump or gap) in the precise single-point positioning assembly
Oracle
FINDING against the published ITRF2020 WTZR coordinate: the pre-registered bars (smoothed median and p95 at most 0.6 x raw; tests/fixtures/gnss_wettzell_2025055/PREREGISTRATION_HATCH.md) were missed, ratios 0.718 and 0.923; the residual metre-level error is bias-like (C1C code bias, antenna, tide, long-period multipath), not the white noise smoothing removes. MODELLED: the filter's statistics are checked internally
ValidatedRTK double differences with LAMBDA integer fixing on real RINEX observationsrtk_real (run_rtk),lambda (resolve)
Capability
Short-baseline relative positioning from real dual-frequency GPS code and carrier phase: double differences against the highest satellite with their full covariance, a float Kalman filter over the baseline and single-difference ambiguities with slip re-initialisation, LAMBDA integer least squares with the ratio test, and the baseline conditioned on the fixed integers
Oracle
Published ITRF2020 coordinates of WTZR and WTZZ (IGN ITRF2020_GNSS.SSC, solution 3, with the RINEX antenna heights) giving the reference baseline; bars pre-registered in tests/fixtures/gnss_wettzell_2025055/PREREGISTRATION_RTK.md before the code was written. Short baseline only: no ionospheric or tropospheric double differences are modelled
Tests
tests/rtk_real_wettzell_oracle.rs (IGS stations WTZR base / WTZZ rover, 1.4 m, 2025-02-24 00-03 h, 30 s: 359 epochs, fixes at 100 % from 00:10, fixed-baseline error median 4.5 mm, p95 5.5 mm; first 10 min fixed 2.6 mm…
ValidatedPPP receiver-clock time solution against the IGS final station clockppp_real (run_ppp,PppEpoch::clock_m)
Capability
The receiver clock estimated by the static float PPP filter on real dual-frequency GPS code and phase with IGS final orbits and clocks: a time transfer from the receiver to IGS time, the PPP iono-free time solution.
Oracle
The IGS final combined clock for station WTZR (AR records of IGS0OPSFIN_20250550000_01D_05M_CLK, SHA-256 d19f3d45...993d, a combination of several analysis centres' network solutions; IGS products, free use with attribution); bars pre-registered in xval/ppp-time/PREREGISTRATION.md before any comparison; all held on the first run. WHAT IS NOT: one station and one two-hour converged window; IGS time, not UTC(k); the constant offset includes receiver code biases and is bounded, not calibrated; no Circular T traceability.
Tests
tests/ppp_time_igs_station_clock.rs (WTZR 2025-02-24, 04:00-05:55, 24 epochs: Kshana minus IGS std 0.075 ns, bar 0.20; mean -0.659 ns, bar 5.0)
ValidatedFloat precise point positioning on real RINEX observationsppp_real (run_ppp,parse_antex_receiver,solid_tide_ecef,phase_windup)
Capability
Static float PPP extended Kalman filter on real dual-frequency GPS code and carrier phase with IGS final orbits and clocks: ionosphere-free combination, satellite and receiver antenna offsets (receiver variations), solid Earth tide, phase wind-up, Shapiro delay, estimated zenith wet delay, per-arc float ambiguities with geometry-free / Melbourne-Wübbena slip detection
Oracle
Published ITRF2020 coordinate of WTZR (IGN ITRF2020_GNSS.SSC, solution 3, propagated with its velocity) — bars pre-registered in tests/fixtures/gnss_wettzell_2025055/PREREGISTRATION_PPP.md before the filter was written; development on the co-located station WTZZ only. Not modelled: ocean loading, satellite PCV, integer ambiguity resolution
Tests
tests/ppp_real_wettzell_oracle.rs (IGS station WTZR, 2025-02-24 00-06 h, 30 s, held out from development: last-epoch horizontal 1.5 mm, up −2.9 cm; 04-06 h 240/240 epochs within 0.10 m 3-D, p95 6.7 cm; code-only p95…
RINEX 4.00-4.02 navigation files: > EPH records of GPS LNAV, Galileo I/NAV and F/NAV, BeiDou D1/D2 and QZSS LNAV decoded into the broadcast-ephemeris type; CNAV, CNV1-3, FDMA, SBAS, STO, EOP and ION records skipped to the next record header; a record whose header names another satellite, or a truncated record, refused
Oracle
The same day's RINEX 3 merged broadcast file of the IGS (published data, written by a different file generator from the same broadcast messages); bars pre-registered in tests/fixtures/brdc_rinex4_2026200/PREREGISTRATION.md (589b7c9) before the first run
Tests
tests/rinex4_nav_brd400_oracle.rs (IGS BRD400DLR_S RINEX 4.02 vs BRDM00DLR_S RINEX 3.04 of 2026-07-19, slice 00-02 h: G 37 / E 478 / C 72 records matched both ways, every field identical, position at Toe 0 m; full day…
RINEX 4.0x observation headers (DOI, LICENSE OF USE, STATION INFORMATION, MERGED FILE; labels matched in columns 61-80 for version 4 and later) and epoch records
Oracle
Format conformance on a real published RINEX 4.00 file (values read back equal the file's printed literals) and a specification-derived header; no independent reader was run, so MODELLED
Tests
tests/rinex4_obs_mao0.rs (first 4 epochs of the IGS MGEX RINEX 4.00 file MAO000USA_R_20231980000_01D_30S_MO, 58 satellites, 7 systems: header records and first-epoch values equal the published literals)…
BDS-SIS-ICD-B1I v3.0 §5.2.4.12 geostationary-satellite position from a broadcast ephemeris (inertial-like frame, then Rz(Ω̇e·tk)·Rx(−5°)), opt-in through the include_beidou_geo field of NavParseOptions, so the default reader still skips GEOs
Oracle
Pre-registered (589b7c9, 3446949) against an IGS MGEX precise orbit with the GEOs; no such product was reachable from the session that built it, so the comparison is pending and the row is MODELLED (checked against the published nominal GEO slots only)
Tests
tests/bds_geo_broadcast_oracle.rs (real C01-C04 records: geostationary radius, longitudes 140.02/80.04/110.55/159.99 deg at the published slots, |latitude| 2.13 deg with the tilt vs 6.65 without; the pre-registered…
Satellite clock offset between the tabulated SP3 clock nodes by linear interpolation over the two bracketing valid nodes (the RTKLIB pephclk scheme; bad-clock sentinel skipped), and the same nodes fed to the precise-product chain as RINEX-clock-equivalent records
Oracle
IGS final 30 s satellite clock combination (IGS0OPSFIN 30S CLK, 2025-02-24): published clock estimates at the query instants, not interpolated from the SP3; bars pre-registered in tests/fixtures/sp3_clock_igs_oracle/PREREGISTRATION.md before the first run (RMS 0.15 ns, p95 0.30 ns, node median 1e-12 s, discrimination 2x)
Tests
tests/sp3_clock_igs_oracle.rs (IGS final 15 min SP3 of 2025-02-24 vs the IGS final 30 s clocks, 32 GPS satellites: node median |Δ| 2.6e-13 s; 3038 mid-point records RMS 0.121 ns, p95 0.265 ns; holding the node instead…
ModelledCarrier-phase tightly coupled GNSS/INS with float ambiguitiesfusion::carrier_tc (run_carrier_tc,CarrierTcCfg),fusion::trajectory (truth_epochs)
Capability
Error-state EKF over 15 INS states, a two-state receiver clock and one float carrier ambiguity per tracked satellite; pseudorange, carrier phase and optional range rate from an SGP4 Walker constellation (light-time and Sagnac), Klobuchar and Saastamoinen/Niell corrections applied in the loop at the filter's own position; time-differenced-carrier slip detection with a per-satellite threshold widened by the velocity uncertainty; ambiguity re-initialisation on slips and re-acquisition
Oracle
Internal consistency: the measurements are synthesised from the models the filter uses (χ² NEES/NIS over seeds, a noise-free convergence, slip injection). MODELLED; no real receiver data or third-party filter is compared
Tests
tests/fusion_carrier_tc.rs (figure-8 drive: steady 3-D RMS 0.025 m carrier vs 0.470 m code on identical measurements; 12 seeds mean NIS/dof 1.00 and steady NEES/dof 0.92; a 1-cycle slip detected at its epoch with no…
ModelledGNSS/INS maritime, rail and aviation packs with a tight-coupling optionfusion::pack (TightCfg,[tight]),scenarios gnss-ins-maritime /gnss-ins-rail /gnss-ins-aviation
Capability
Three gnss-ins scenarios on the trajectory library (a vessel weaving at 8 m/s with a five-minute outage; a train through a two-satellite cutting and a tunnel; an aircraft whose outage covers a 25-degree banked turn) and the pack's optional [tight] table: pseudoranges from an SGP4 Walker constellation fused by the closed-loop navigator, every satellite above the mask when nominal and the highest few when degraded
Oracle
Internal: each pack's figures of merit against its own truth and a loose/tight comparison on identical inputs. MODELLED; the audit's NaveGo comparison has not been run
Tests
tests/gnss_ins_packs.rs (each pack: fused below free INS over the outage for both sensors; rail: tight 5.3 m vs loose 12.9 m navigation-grade and 15.6 m vs 18.8 m tactical over the outage, availability 0.97 vs 0.95…
ModelledCoupled pseudorange PNT filter with bias state and 3-D extension in the fusion packfusion::coupled (CoupledPntFilter,CoupledPntBiasFilter,CoupledPnt3d),fusion (run_coupled_suite),hybrid (CoupledCfg)
Capability
One stacked Kalman filter over position, velocity, clock phase and frequency fed pseudoranges that couple position and clock (Joseph-form updates carrying the cross-covariance); a 1-DOF variant augmented with a constant accelerometer-bias state that the fusion pack runs under [coupled]; and the 3-D 8-state version with line-of-sight rows
Oracle
Internal consistency: truth simulated from the filters' own models (χ² NEES over Monte Carlo seeds, the decoupled pack on identical sensors). MODELLED
Tests
tests/fusion_coupled_filters.rs (3-D filter over 60 seeds: second-half mean NEES 8.12 for 8 states, last-epoch χ² sum inside its 95 % band; a 0.1 ns clock fix cuts the 3-D position sigma from 5.58 m to 3.07 m; bias…
ValidatedLoosely coupled GNSS/INS EKF with lever arm on the KF-GINS datasetfusion::kfgins (run,loaders,score),fusion::closed_loop (propagate_increments,fuse_position_lever_arm),fusion::gnss_ins_ekf (update_position_lever_arm)
Capability
The closed-loop 15-state error-state EKF driven by IMU increments with GNSS antenna positions fused through the lever-arm model H = [I 0 −[(C_b^n l_b)×] 0 0], configured from a KF-GINS YAML file (initial state and sigmas, ARW/VRW, Gauss-Markov biases)
Oracle
The dataset's published reference trajectory and the KF-GINS C++ program (GPL-3.0, commit 8291a93e, run as a separate tool); bars pre-registered in xval/kfgins/PREREGISTRATION.md, and KF-GINS was not run on this data before that registration (checked in its Results). All six bars held on the first run. DATA-GATED: the KF-GINS demo dataset (real Leador-A15 IMU and RTK GNSS with a reference trajectory) carries no licence, so it is not vendored; xval/kfgins/fetch.sh fetches it at test time with every file's SHA-256 pinned, and the test fails on a mismatch; the reproduce path, which needs network access, is in xval/kfgins/PREREGISTRATION.md. Promoted to VALIDATED by the owner's decision on data-gated evidence (2026-10-13)
Tests
tests/fusion_kfgins_dataset_oracle.rs (synthetic loaders and spinning-platform lever-arm tests always; data-gated on KSHANA_KFGINS_DIR, run 2026-10-13 on the KF-GINS demo dataset: horizontal / vertical RMS 0.0109 /…
ValidatedTest-bench export read by a simulator's motion-file readers (gps-sdr-sim)interop::testbench (write_motion_csv,write_nmea,export)
Capability
The bench-export motion files given to an open-source GNSS simulator's own readers come back as the exported trajectory: the NMEA GGA file (-g), the time/latitude/longitude/height cut of the motion CSV (-x) and its time/x/y/z cut (-u), on 6 gnss-ins scenarios (including the maritime, rail and aviation packs, read to the simulator's 3000-sample cap) and 2 stationary ones. Motion only; no signal is written or generated.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: gps-sdr-sim (MIT, commit 28ca29a6) readNmeaGGA, readUserMotion and readUserMotionLLH with its own WGS 84 llh2xyz, called through xval/testbench-gpssdrsim/driver.c (no signal function called), against the motion CSV's Earth-fixed columns row by row, with bars pre-registered in xval/testbench-gpssdrsim/PREREGISTRATION.md before any run: every sample read on all 8 scenarios; NMEA worst 1.21 mm (bar 1 cm); LLH cut worst 0.145 mm (bar 1 mm); four-column Earth-fixed cut exact (bar 1 um, amendment 1); gnss-ins rows on the simulator's 10 Hz grid to 2.3e-13 s. FINDING (run 1, pinned): the header-removed full-width CSV is not readable by -u, because the reader's 100-character line buffer splits each 161-character row into two samples; docs/TEST-BENCH.md gives the four-column cut. WHAT IS NOT: the physics of the scenario's truth trajectory; gps-sdr-sim's signal output; any other simulator's reader; any receiver.
Tests
tests/testbench_gpssdrsim_oracle.rs (6 tests: oracle and scenarios, export bytes pinned, NMEA and LLH readers, four-column ECEF cut, run-1 FINDING on the full-width CSV, 10 Hz row grid)
ModelledGNSS/INS trajectory library and CSV truth importfusion::trajectory (TrajectoryCfg,Trajectory,kinematic_imu,parse_truth_csv),fusion::pack ([trajectory])
Capability
Selectable motion profiles for the gnss-ins pack (straight, figure-8, urban stop-and-go, aircraft climb and coordinated turn, survey drone), defined by Euler-rate and along-track-acceleration commands closed on the truth state with the kinematic IMU (centripetal, gravity and Coriolis terms), and imported CSV truth turned into the IMU that reproduces it; the historical drive unchanged when no profile is given
Oracle
Internal: rigid-body kinematics identities, the CSV round trip and the filter's own covariance. MODELLED (no real drive log is compared)
Tests
fusion::trajectory::tests (coordinated turn has no side force and load factor 1/cos φ; CSV import reproduces a flown profile to 1e-3 m/s and 1 m); tests/gnss_ins_trajectory_library.rs (every profile runs and fusion…
ValidatedStrapdown INS mechanizationinertial::mechanization,inertial::attitude,inertial::imu_errors
Capability
Quaternion-attitude, WGS-84 NED strapdown inertial mechanization (coning/sculling-compensated) propagating a navigation state from (Δθ, Δv) increments
Oracle
NaveGo v1.4 (R. Gonzalez et al., LGPL-3) run under Octave — an independent published INS toolbox driven by the identical (Δθ,Δv) increment stream. Attitude matches bit-for-bit (same NED / scalar-first-quaternion / Earth-rate / transport-rate conventions); velocity/position agree to two documented differences — the deflection-of-vertical north-gravity term NaveGo includes and kshana omits by design (plumb-bob gravity; matched to the Groves closed form to every digit, with a sanity-floor assert) and O(dt²) integrator differences — not mechanization errors
Tests
tests/classical_strapdown_ins_reference.rs (static/turn/coning profiles, 30 epochs each, vs NaveGo: attitude bit-identical 0 rad; velocity/position within named analytic bounds)
ValidatedGravity-field functional synthesis (gravity-aided / GNSS-free nav map)gravity_sh
Capability
Spherical-harmonic gravity magnitude + disturbance (mGal) from any ICGEM .gfc model; GRS80 normal gravity
Oracle
GRS80 (Moritz 1980, IAG) Somigliana normal gravity + published γ_e/γ_p; real ICGEM EGM2008 field
Tests
tests/icgem_gravity_reference.rs (GRS80 synthesis reproduces Somigliana to 3.5e-12; real ICGEM EGM2008 disturbance map physical)
ModelledEarth-Moon CR3BP 9:2 NRHO against the orbit CAPSTONE flewcr3bp (differential_correct_halo,PeriodicOrbit)
Capability
The circular-restricted 9:2 L2 southern near-rectilinear halo orbit that cr3bp::differential_correct_halo corrects (the Gateway and CAPSTONE orbit family), compared for its mean geometry (period, perilune radius, apolune radius) with the orbit CAPSTONE actually flew. A measurement of how well the CR3BP member describes the real orbit; a CR3BP orbit is not an ephemeris-model orbit and is not compared epoch by epoch
Oracle
Measured against the JPL Horizons reconstructed CAPSTONE ephemeris (NAIF -1176, CAPSTONE_merged, Moon-centred ICRF, 2023-03-01..31 at 10 min), committed byte for byte (tests/fixtures/capstone_horizons/, NOTICE.md), pre-registered dab3215c before it was fetched, as a measurement with bars on the mean geometry. Not a validation: the CR3BP omits the Sun, the Moon's eccentricity and the station keeping CAPSTONE performed, so the bars describe how representative the model orbit is, not an accuracy. Results: docs/design/evidence/cap29d-capstone-nrho/RESULTS.md
Tests
tests/capstone_nrho_horizons.rs::cr3bp_nrho_matches_capstone_mean_geometry (CAPSTONE March 2023, 4 perilunes and 5 apolunes: period 6.5004 d vs 6.5699 d (1.07 %, bar 3 %), apolune 71 182 vs 71 215 km (0.05 %, bar 5 %)…
ValidatedCR3BP rotating states mapped to Moon-centred ICRF on the DE440 Earth-Moon linecr3bp (to_moon_centred_icrf,from_moon_centred_icrf,to_moon_centred_icrf_de440)
Capability
cr3bp::to_moon_centred_icrf maps a nondimensional Earth-Moon CR3BP rotating state to a Moon-centred ICRF position and velocity on the instantaneous Earth-Moon line (length unit the Earth-Moon distance, time unit the inverse orbital rate, the frame's full angular velocity including its out-of-plane rate from the Moon's acceleration), with the inverse from_moon_centred_icrf, and cr3bp::to_moon_centred_icrf_de440 reads the geocentric Moon state from a DE440 SPK. It maps states; it does not make a CR3BP orbit an ephemeris-model orbit
Oracle
NAIF SPICE Toolkit (CSPICE N0067 through spiceypy 8.3.0) sxform of a two-vector dynamic frame (primary +X along the Moon's position wrt the Earth, secondary +Y along its velocity, de440s), applied to the dimensional rotating state built from SPICE's own Moon state, by a separate program that calls no Kshana code (tests/fixtures/cr3bp_to_mci_spice/, NOTICE.md). Pre-registered 3bfc9445, amended before any run in e8bfabe4 to include the frame's out-of-plane angular velocity (without it the velocity would be up to 8e-2 m/s off). A Library check of the transform only. Results: docs/design/evidence/rm34-cr3bp-to-mci/RESULTS.md
Tests
tests/cr3bp_to_mci_spice_oracle.rs::cr3bp_to_mci_matches_a_spice_two_vector_frame (72 cases: 12 epochs in January 2025 x the L1 and L2 points, two L2 southern halo and one DRO initial state of the JPL Three-Body…
ValidatedSolar-system kind on a JPL kernel (planetary_kernel_path)solar_system (SolarSystemScenario::planetary_kernel_path,KernelOut); ephem_provider (KernelEphemeris)
Capability
The solar-system kind with planetary_kernel_path: each reported body whose own NAIF code the SPK holds takes its heliocentric J2000 position and velocity from the kernel through the engine's own reader (label KERNEL), a link whose two ends are both in the kernel is computed on it, barycentres are never substituted for a planet, and every other body stays analytic with its own label; the report gains a kernel block (path, SHA-256, the bodies and links taken from it) only when the option is set. With de440s.bsp that is Mercury, Venus, the Earth and the Moon
Oracle
NAIF SPICE Toolkit (CSPICE N0067 through spiceypy 8.3.0) spkezr heliocentric states on the full de440s.bsp, a separate program that calls no Kshana code (tests/fixtures/solar_system_kernel/, NOTICE.md); the engine reads bit-for-bit de440s cuts. Pre-registered f4a6297a. A Library check of the wiring: DE440 itself and the analytic bodies are not validated here. Results: docs/design/evidence/cap31-solar-system-kernel/RESULTS.md
Tests
tests/solar_system_kernel_oracle.rs::kernel_backed_solar_system_matches_spice (Mercury, Venus, the Earth and the Moon at 2026-01-01, 2030-06-15 and 2040-12-31: worst 3.3e-5 m and 1.0e-11 m/s against 1 mm and 1e-6 m/s…
ValidatedPorkchop on the DE440 ephemerismaneuver (porkchop_kernel,PorkchopGrid); ephem_provider (KernelEphemeris)
Capability
maneuver::porkchop_kernel: the launch-date x arrival-date sweep of departure C3 (km²/s²) and arrival V-infinity (km/s) with the single-revolution prograde Lambert solver and the 0.5 deg collinearity guard, on the heliocentric J2000 states of any two NAIF bodies read from a JPL DE440 SPK by the crate's own reader. One revolution, prograde, no launch-vehicle or arrival constraints: a sweep, not a mission design
Oracle
NAIF SPICE Toolkit (CSPICE N0067 through spiceypy 8.3.0) heliocentric states on de440s.bsp and pykep 3.0.1 lambert_problem (ESA, the C++ kep3 core, MPL-2.0), an independent Lambert solver, run as a separate program that calls no Kshana code (tests/fixtures/porkchop_de440/, NOTICE.md); the engine reads a bit-for-bit de440s cut. Pre-registered b9317cee. A Library check of the sweep. Results: docs/design/evidence/cap20b-porkchop-de440/RESULTS.md
Tests
tests/porkchop_de440_oracle.rs::de440_porkchop_matches_spice_and_pykep (Earth to the Mars barycentre, 16 departures 2026-09-01..2027-01-29 x 19 arrivals 2027-06-01..2028-02-26: 304 cells finite on both sides, worst C3…
ValidatedImpulsive transfer optimiser on the DE440 ephemerismaneuver (optimise_transfer_kernel,TransferObjective,TransferType,TransferOptimum); ephem_provider (KernelEphemeris)
Capability
maneuver::optimise_transfer_kernel: the departure and arrival dates of a single-revolution ballistic two-impulse transfer that minimise departure C3 or arrival V-infinity within Type I or Type II, as a continuous optimum (a 2-day grid over both windows, then 2-D Nelder-Mead), on heliocentric states from the DE440 SPK. A date optimiser for one leg, not a mission design: no launch-vehicle, parking-orbit, deep-space-manoeuvre or multi-revolution options
Oracle
An independent optimiser: NAIF SPICE Toolkit (CSPICE N0067 through spiceypy 8.3.0) heliocentric states on the same de440s cut, pykep 3.0.1 lambert_problem (ESA, MPL-2.0) and SciPy 1.18.1 Nelder-Mead, calling no Kshana code. The NASA Earth-to-Mars handbook (MIDAS, ephemeris unstated, 2-day date grid) is a secondary published reference and does not set the status
Tests
tests/transfer_optimiser_oracle.rs::transfer_optimiser_matches_spice_pykep_scipy (Earth to the Mars barycentre, departures 2026-09-15..2027-01-15 x arrivals 2027-05-01..2028-03-01, Type I/II x min C3 / min V-infinity…
ValidatedMulti-revolution Lambert transfer solvermaneuver (lambert_multirev,LambertSolution,LambertBranch)
Capability
Every solution of Lambert's problem with up to three complete revolutions (maneuver::lambert_multirev: the single-revolution solution, then the left and right branches of each revolution count, Izzo 2015), with the boundary velocities v1 and v2, for prograde and retrograde transfers, in canonical and heliocentric SI units, including transfers within 1 % of a revolution count's minimum time of flight and near-180 deg geometries. A two-body solver only (no perturbations)
Oracle
pykep 3.0.1 lambert_problem (ESA, the C++ kep3 core, MPL-2.0), an independent library, run as a separate program by scripts/gen_lambert_multirev_ref.py; only its numerical output is committed. Pre-registered 9879ae57 (case set, bars, outcome rule) before any comparison; the reference was regenerated independently and is byte-identical. Results: docs/design/evidence/lambert-multirev/RESULTS.md
Tests
tests/lambert_multirev_oracle.rs::multirev_lambert_matches_pykep_on_the_preregistered_cases (382 cases: 300 random 3-D, 18 near 180 deg, 24 near the minimum time of M = 1..3, 40 heliocentric; 1058 solutions; counts and…
ValidatedLambert two-body transfer solvermaneuver (lambert)
Capability
Izzo-2015 single-revolution Lambert solver (r1, r2, time-of-flight → boundary velocities) across LEO→GEO→heliocentric transfers, prograde and retrograde
Oracle
lamberthub 1.0.0 izzo2015 (J. Martínez Garrido, MIT) — an independent third-party Lambert solver. The single-revolution (M=0) Lambert problem has a unique solution, so library-vs-library agreement is a genuine external check; matched to <1e-4 m/s (observed ~1e-11), the same kind of validation DOP gets vs gnss_lib_py
Tests
tests/lambert_reference.rs (13 transfers vs lamberthub izzo2015; worst |Δv| 7e-12 m/s)
ValidatedTwo-part Julian dates and leap seconds on the Earth-orbit time pathjd2 (Jd2::from_utc_calendar,utc_to_tai,tai_to_utc,utc_to_ut1,tai_to_tt,earth_rotation_angle); sgp4 (Sgp4::new_at,Sgp4::propagate_at)
Capability
A UTC calendar date and time to a two-part UTC Julian date with the SOFA dtf2d convention for days that end in a leap second (86 401 s long, 23:59:60.x valid), UTC to TAI and back, and UT1 from UTC through TAI, on two-part dates from 1972-01-01: the time path SGP4 (Sgp4::propagate_at, measured from a two-part epoch) and the IAU 2006/2000A Earth-fixed rotation (sgp4::teme_to_itrs_matrix_eop with the two-part Earth rotation angle) take
Oracle
pyerfa 2.0.1.5 (liberfa 2.0.1, BSD-3-Clause; the SOFA algorithms as released by the ERFA project), an independent library: erfa.dtf2d("UTC"), erfa.utctai and erfa.taiutc on the committed instants, the gap formed part by part on the two-part dates, bar 1e-6 s fixed before the run (pre-registered 896f3c2e). Worst gap 9.6e-12 s (dtf2d) and 1.9e-11 s (utctai, taiutc). Every UTC day taken as 86 400 s, or the leap-day fraction left unscaled, turns the test red. UT1 through TAI (iauUtcut1) is checked by the Earth-orbit path row's ERFA leg. Scope: UTC from 1972-01-01; the earlier drifting-rate UTC is not modelled
Tests
tests/jd2_sofa_time_oracle.rs (224 UTC instants: six around each of the 27 leap seconds 1972-2016, two fixed, 60 seeded across 1972-2026; dtf2d, utctai and taiutc each within 1e-6 s, worst 1.9e-11 s); jd2::tests…
ValidatedTwo-part Julian dates through Earth rotation on the orbit pathjd2 (gmst82,From<TwoPartJd>,From<Jd2>); cio (gcrs_to_itrs_matrix_jd2); frames (teme_to_itrf_jd2,itrf_to_teme_jd2); orbit (Propagator::position_in_frame_jd2,own_jd2_utc); ephemeris (EphemerisScenario::two_part_jd)
Capability
Earth rotation from a two-part UT1 Julian date on the orbit path, so the Earth-fixed position keeps the microsecond a single-f64 date loses (about 3e-9 rad near 2020): jd2::gmst82 (IAU 1982 GMST, the TEME convention, on the two parts), cio::gcrs_to_itrs_matrix_jd2 (the CIO GCRS to ITRS matrix with the Earth rotation angle from the two parts), and, built on them and checked internally, frames::teme_to_itrf_jd2 / itrf_to_teme_jd2, orbit::Propagator::position_in_frame_jd2 and own_jd2_utc, the ephemeris scenario's opt-in two_part_jd (the ECEF position, ground track and station view; default off, unchanged), and lossless From conversions between jd2::Jd2 and timescales::TwoPartJd. Precession-nutation and the Cowell force-model epoch stay single-f64, where a 40 us rounding is below 1e-15 rad or far below the analytic Sun/Moon ephemeris error
Oracle
pyerfa 2.0.1.5 (liberfa 2.0.1, BSD-3-Clause; the SOFA algorithms as released by the ERFA project), an independent library: gmst82 and c2t06a on the committed two-part dates, polar motion 0.1 and 0.3 arcsec. Pre-registered 765860c3 (dates, bars) before the functions were written and before the oracle was run. The internal tests (the scenario option, the TEME/ITRF round trip, the type conversions) are supporting self-consistency
Tests
tests/jd2_earth_rotation_erfa_oracle.rs::two_part_earth_rotation_agrees_with_erfa (40 two-part UT1/TT dates 1980-2050 including +1, +10, +50 us steps at two dates: gmst82 within 1e-12 rad, worst 0; GCRS to ITRS…
ValidatedFrame matrices against ERFA vectors: IAU 2006 bias-precession, TEME to GCRS, and the CIO chain with celestial-pole offsetsprecession (precession_matrix); nutation (teme_to_gcrs_matrix); cio (gcrs_to_cirs_matrix_with_offsets)
Capability
Three frame matrices compared element by element or by rotation angle with ERFA at arbitrary epochs: precession::precession_matrix (IAU 2006 bias-precession by Fukushima-Williams angles), nutation::teme_to_gcrs_matrix (the TEME to GCRS rotation P^T N^T Rz(-EE) with IAU 2006 precession, IAU 2000B nutation and the equation of the equinoxes with its two leading complementary terms), and cio::gcrs_to_cirs_matrix_with_offsets (the GCRS to CIRS rotation with the IERS celestial-pole offsets dX, dY added to the model CIP X, Y before the rotation, the SOFA cookbook convention; new). The model choices themselves (2000B nutation on the TEME path, the two-term complementary series) are not validated, only their implementation
Oracle
pyerfa 2.0.1.5 (liberfa 2.0.1, BSD-3-Clause; the SOFA algorithms as released by the ERFA project), an independent library: pmat06; numat(obl06, nut00b) with the equation of the equinoxes from eect00 (the full complementary series), composed as P^T N^T Rz(-EE); xys06a then c2ixys(X + dX, Y + dY, s). Pre-registered bbaeb5b5 (epochs, offsets and bars) before the oracle was run. The 1.1e-10 rad TEME residual is the predicted effect of keeping two complementary terms (about 3 mm at GNSS orbit)
Tests
tests/frames_erfa_vector_oracle.rs::frame_matrices_agree_with_erfa_vectors (40 TT epochs 1950-2100, dX and dY up to 0.5 mas: precession elements within 1e-13, worst 0; TEME to GCRS within 5e-10 rad, worst 1.10e-10 rad…
ValidatedPre-1972 UTC (drifting-rate era) on the two-part Julian-date time pathtimescales (tai_minus_utc_on_date); jd2 (Jd2::from_utc_calendar,utc_to_tai,tai_to_utc)
Capability
UTC from 1960-01-01 to 1971-12-31, when UTC ran at an offset rate from TAI and stepped by fractions of a second, on the two-part Julian-date path: TAI - UTC from the drift table of each period (timescales::tai_minus_utc_on_date), the UTC calendar date to a two-part UTC Julian date with the day length 86 400 s + dleap (Jd2::from_utc_calendar), UTC to TAI with the day fraction scaled from UTC to SI seconds (jd2::utc_to_tai) and TAI back to UTC (jd2::tai_to_utc), following SOFA dtf2d, utctai and taiutc. Before 1960 TAI - UTC is taken as 0 (SOFA's value; UTC undefined) and is outside the claim. The single-f64 timescales::tai_minus_utc still clamps to 10 s before 1972 and is not covered
Oracle
pyerfa 2.0.1.5 (liberfa 2.0.1, BSD-3-Clause; the SOFA algorithms as released by the ERFA project), an independent library: erfa.dtf2d("UTC"), erfa.utctai and erfa.taiutc on the committed instants, the gap formed part by part. Pre-registered 247a01a6 (instants, tolerance and the 1972+ checksum) before the engine change and before the oracle was run. The drift table is transcribed from the same SOFA iauDat table the oracle uses, so the comparison covers the conversion arithmetic (day length, UTC-to-SI scaling, iteration), not the historical table itself
Tests
tests/jd2_pre1972_sofa_oracle.rs::pre_1972_utc_conversions_agree_with_erfa (117 UTC instants: four around each of the 15 starts of a TAI-UTC expression 1960-1972 and 60 seeded across 1960-1971; dtf2d, utctai and taiutc…
ValidatedGLONASS broadcast satellite clock and broadcast-orbit propagationglonass (GlonassEphemeris::sv_clock_bias_s,position_ecef,parse_glonass_nav)
Capability
The GLONASS broadcast satellite clock bias -tau_n + gamma_n (t - t_b) at a time read on the satellite clock (glonass::GlonassEphemeris::sv_clock_bias_s, the offset corrected for the clock bias by two fixed-point iterations), and the PZ-90 Earth-fixed position of the broadcast state vector propagated by RK4 through the GLONASS ICD equations of motion within +/-15 min of the reference epoch (GlonassEphemeris::position_ecef), from real RINEX 3 GLONASS navigation records. No PZ-90.11 to ITRF transformation is applied (they agree at the centimetre level since 2014; RTKLIB applies none either). Not wired into a scenario output: library function
Oracle
RTKLIB v2.4.2-p13 (T. Takasu, BSD-2-Clause, commit 71db0ffa) geph2clk and geph2pos, compiled from C source in a separate harness (xval/rtklib-glonass) that reads the same RINEX file with RTKLIB's own reader, so the parse, the sign of -tau_n, the units and the iteration are compared as well as the clock and the orbit. Pre-registered 5cad23b7 before the function existed and before the harness was built. Both sides evaluate the same clock closed form, so the clock comparison covers the implementation, not the formula. The orbit integrators differ (RTKLIB steps 60 s then the remainder, Kshana equal steps of at most 60 s) and so do the constants (RTKLIB mu 3.9860044e14, Kshana the PZ-90.11 3.986004418e14); the 1.1 mm worst case is that difference, as predicted before the run. Scope: the broadcast ephemeris within its validity interval, not the true orbit or clock (no comparison with precise products)
Tests
tests/glonass_broadcast_rtklib_oracle.rs::glonass_broadcast_clock_and_orbit_match_rtklib (292 cases: the 4 GLONASS records of the real 2013-01-01 IGS MGEX broadcast file x 73 offsets within +/-900 s; clock within 1e-14…
ValidatedOne Earth-orbit path: SGP4 element sets to the GCRS and ITRS on the IAU 2006/2000A chain with Earth orientation parameterssgp4 (MeanElementSet,SgpOrbit,Eop,teme_to_itrs_matrix_eop,gcrs_to_itrs_matrix_eop); jd2 (utc_to_ut1,earth_rotation_angle); nutation (teme_to_gcrs_matrix); cio (gcrs_to_cirs_matrix); frames (polar_motion_matrix)
Capability
The propagation and frame path every routed Earth-orbit kind shares: an SGP4 mean element set with a two-part UTC epoch (sgp4::MeanElementSet, read from constellation-design elements by MeanElementSet::from_earth_fixed) propagated by SGP4/SDP4 to TEME (sgp4::SgpOrbit), and the TEME -> ITRS and GCRS -> ITRS rotations of the IAU 2006/2000A chain with UT1 - UTC and polar motion (sgp4::teme_to_itrs_matrix_eop, gcrs_to_itrs_matrix_eop, the zero-parameter forms teme_to_itrs_matrix and gcrs_to_itrs_matrix). TEME is the true equator of date with the mean equinox, reached from the GCRS through the IAU 2006 bias-precession, the IAU 2000B nutation and the equation of the equinoxes
Oracle
Leg 1: python-sgp4 2.24 (MIT licence; D. Vallado's reference C++ SGP4, the code that produced the AIAA 2006-6753 verification vectors), WGS-72, improved mode. Leg 2: pyerfa 2.0.1.5 (ERFA, BSD-3-Clause, the SOFA routines): c2t06a with TT from utctai/taitt and UT1 from utcut1, and for TEME c2t06a . pnm06a^T . rz(-ee06a). Bars fixed before the run (pre-registered d92aef9b). Disclosed: the first run failed leg 2 at 2016-12-31T23:59:59.5 by one second of rotation (UT1 taken from the UTC quasi Julian date on a leap-second day); the engine was corrected (jd2::utc_to_ut1) and the second run agrees at the unchanged bars. Polar motion dropped (46 failures) or WGS-84 constants in place of WGS-72 (4432 failures) turn the test red. This three-leg design replaced an end-to-end Orekit comparison that failed on Orekit's SDP4 at e = 0 and on its IERS 1996 TEME (the finding stays pinned in tests/leo_polar_coverage_full_claim_orekit_oracle.rs)
Tests
tests/earth_orbit_path_sgp4_erfa_oracle.rs (leg 1: 4432 TEME positions of the 232 committed element sets against python-sgp4 2.24, worst 6.1e-8 m, bar 1 cm; leg 2: 108 instants, 23 of them with Earth orientation…
ModelledPosition error from geometry: RMS error = sigma_UERE x DOP for the engine's least-squares fixdop_error_mc (run_geometry,check,TOLERANCES); dop_error_ci (run,GEOMETRIES); pvt (solve_spp); orbit (dop,dop_cofactor)
Capability
dop_error_mc::run_geometry and check: a seeded Monte Carlo (N = 20,000 epochs per geometry, sigma = 1.5 m, ChaCha8 seed 36 + i) of the engine's own iterated least-squares fix (pvt::solve_spp, unit weights, solved from the true position) on the 8 geometries of tests/fixtures/dop/dop_reference.csv, comparing the east, north, up and clock error variances, the H/V/P RMS, the east-north correlation, the mean errors and the fraction inside the 95 % horizontal ellipse with sigma^2 Q, where Q is the covariance factor of the orbit::dop kernel (orbit::dop_cofactor, the same arithmetic) on the same geometry
Oracle
Internal consistency against a closed form: under the linear Gaussian model the least-squares covariance is sigma^2 (H^T H)^-1 (Kaplan and Hegarty (eds.), Understanding GPS: Principles and Applications, 2nd ed., Artech House, 2006, ch. 7). Method and tolerances of the 0.36 geometry design note, section 5.2, fixed in a pre-registration commit before the first run: each variance within 5 % (5 standard errors of a sample variance at N = 20,000), each RMS within 2.5 %, the east-north correlation within 0.035 (5/sqrt(N)), each mean within 5 standard errors, the 95 % ellipse fraction within 0.95 +/- 0.0077 (5 binomial standard errors); negative controls (PDOP doubled; errors growing as 1/sin(elevation) against the unweighted Q) must fail. HOW IT COMPOSES: Q is the covariance factor of the orbit::dop kernel, VALIDATED against gnss_lib_py 1.0.4 on these same 8 geometries in tests/dop_reference.rs (row 'GNSS geometry / dilution of precision'); this row adds the step from those factors to metres through the engine's solver and does not borrow that row's status. WHY MODELLED: no external dataset is involved; agreement shows the estimator and noise realise sigma x DOP under the textbook model, not that any receiver's error equals UERE x DOP, and sigma_UERE is an input
Tests
tests/uere_dop_monte_carlo.rs (position_error_is_sigma_times_dop_on_the_gnss_lib_py_geometries: all 104 checks pass on the first run after pre-registration f97bf4c1, worst variance 3.2 %, RMS 1.6 %, correlation 0.012…
ModelledGeometry-aware trust classification: satellite geometry beside the receiver-trust score and a class on every deductionreceiver_trust (geometry: GeometryTracker,geometry_ratio,sky_ratio,geometry_points,GeometryLabel; score: DeductionClass,score_with; monitors; live); telemetry (sample,prometheus,syslog,otlp)
Capability
receiver_trust::geometry: per-epoch satellite geometry (good, degraded with reasons few-satellites/high-pdop/high-hdop/sky-blocked, or unknown when the log gives no GSV, NAV-SAT or GSA) derived from the reported elevations and azimuths through orbit::dop or taken from GSA, with a label always reported and never scored (good, fair, poor, unknown) and the HDOP, PDOP, VDOP, satellite count and signal loss in words; a vessel-score geometry monitor that deducts nothing unless the geometry is poor (ratio max(hdop/4, pdop/6, 5/n_used) above 1), then ramps to a cap of 15 points at 1.5 that geometry alone cannot take below the degraded band (cap, ramp end and floor in [geometry]; no alarm); a class on every deduction (geometry, environment, inconsistency) with loss-of-lock to satellites below 15 deg classed geometry, satellites in view but not tracked classed environment, and a poor-geometry deduction whose stated HDOP is beyond max(1.0, 25 % of the derived HDOP) from the stated sky classed inconsistency
Oracle
Internal consistency: the classification is checked against the rules it states and against hand-derived ratios on synthetic edits of the committed demo log, and the no-change pins hold the 0.35.0 outputs: no score changes unless the geometry is poor; the label is always reported. WHY MODELLED: the thresholds are chosen, not fitted (provenance in each run's geometry_model; the IEC 61108-1 and GPS SPS figures as cited, to be checked against the standards' text), and there is no external labelled dataset of natural geometry degradation against attacks to validate the classification against. The derived DOP reuses the orbit::dop kernel, VALIDATED against gnss_lib_py, whose status this row does not borrow. A class of geometry is a statement about the sky, not a finding that no attack occurred
Tests
tests/natural_degradation_trust.rs (a_stated_skyline_reads_the_fjord_as_geometry: on the synthetic fjord passage of scenarios/training/fjord-natural-degradation.toml the same 624 epochs are deducted with and without…
ValidatedBroadcast ionosphere model (Klobuchar, IS-GPS-200)gnss_sim (klobuchar_delay_m)
Capability
Klobuchar single-frequency L1 slant ionospheric group delay from the eight broadcast α/β coefficients
Oracle
RTKLIB ionmodel (tomojitakasu/RTKLIB, src/rtkcmn.c) — independent reference implementation compiled from source; matched to < 1e-4 m
Tests
tests/klobuchar_reference.rs (10 cases across elevation/azimuth/local-time, two coefficient sets)
The distributional core every protection level rests on: the snapshot fault-detection threshold χ²₁₋ₚfₐ(dof), the missed-detection non-centrality pbias=√λ, and the K_fa/K_md/K_V solution-separation multipliers
Oracle
SciPy 1.17.0 (scipy.stats.chi2/.norm/.ncx2 + optimize.brentq) — independent library (Cephes/Boost), a different algorithm from Kshana's incomplete-gamma series; matched to ≤1e-6 rel. Kernel only. The ARAIM MHSS P_HMI budget *allocation* is no longer without a published numeric oracle — the WG-C ARAIM Technical Subgroup's own worked example now backs the separate 'ARAIM MHSS protection levels against published reference vectors' row, matched at the reference's own TOL_PL = 5e-2 m — so this row's scope is the statistical kernel and that row carries the allocation (see docs/ARAIM_REFERENCE.md)
Tests
tests/raim_reference.rs (171 cases: χ² CDF/quantile, normal CDF/quantile, non-central χ² CDF, pbias across the P_fa/P_md/redundancy ranges)
DO-229E Appendix J weighted-least-squares protection levels: D=(GᵀWG)⁻¹ from per-satellite elevation/azimuth and error budget, horizontal error-ellipse major axis and vertical σ, scaled by the published K-factors
Oracle
RTKLIB SBAS-PL fork — zsiki/rtklib_ws waasprotlevels() (Siki & Takács 2017, "DO-229D Appendix J"), run by rnx2rtkp -ws on real EGNOS GEO-PRN120 messages + real BUTE/Budapest RINEX; independent third-party implementation, HPL matched to < 2e-3 m. gLAB v6.0.0 (core/filter.c) confirmed identical convention. Both oracles round K_V→5.33 vs Kshana's exact Φ⁻¹(1−5e-8)=5.3267 (~0.06%), so the vertical is checked as the K-factor-free d_U
ValidatedAnomaly-detection scoring on real spacecraft telemetryimpairment_eval,eval_stats
Capability
ROC AUC + bootstrap CI separating real labelled anomalies; transparent detector (reproduces-labels)
Oracle
scikit-learn roc_auc_score on the OPSSAT-AD test split (Ruszczak et al. 2025, CC BY 4.0) — real OPS-SAT telemetry
Tests
tests/opssat_ad_reference.rs (real ESA OPS-SAT, AUC reproduces scikit-learn to 1e-9); tests/ai_ml_rf_impairment_detection_evaluation_reference.rs (122 cases on the real OPSSAT-AD test split: full operating-point…
tests/scipy_reference.rs (NNLS; χ² at operating dof ≥ 48; van-Loan Q)
ValidatedGeomagnetic reference field (IGRF-14 synthesis)igrf,igrf_data
Capability
Spherical-harmonic synthesis of the IGRF-14 main field — X/Y/Z/F components, declination and inclination — feeding the magnetic-anomaly alt-PNT layer
Oracle
ppigrf 2.1.0 (K. M. Laundal, MIT) — the IAGA-VMOD pure-Python IGRF reference implementation shipping the official IGRF14.shc coefficients (IAGA 14th generation, Zenodo 10.5281/zenodo.14012302); an independent third-party codebase computing the uniquely-defined IGRF-14 field, matched over a global grid
Tests
tests/alternative_complementary_pnt_reference.rs (2520 global points × altitudes vs ppigrf @ epoch 2025.0; worst |ΔXYZF| 3.9e-4 nT, |ΔD| 2.8e-6°, |ΔI| 3.6e-7°)
Analytic binormal ROC AUC = Φ(μ/(σ√2)) and the minimum detectable fault σ·(Φ⁻¹(1−P_fa)+Φ⁻¹(P_d)) underpinning the quantum-fault and anomaly detectors
Oracle
scipy 1.17 (Cephes ndtr/ndtri) + scikit-learn roc_auc_score (Pedregosa et al., JMLR 2011), both BSD-3-Clause — independent libraries computing the same uniquely-defined Gaussian-tail / AUC quantities, matched to the A&S-erf floor (~7e-8). The quantum-vs-classical advantage built on top stays MODELLED
Rank-correlation and resampling kernels under the resilience decision-instability study: Kendall τ-b, Dirichlet mean, competition ranking and percentile confidence intervals
Oracle
scipy 1.18 (stats.kendalltau variant='b', rankdata) + numpy.percentile (BSD-3-Clause) — independent implementations (merge-sort τ) of the uniquely-defined rank statistics, matched to 1e-12. The decision-instability study built on these kernels stays MODELLED
Tests
tests/resilience_score_decision_instability_reference.rs (124 cases vs scipy 1.18 / numpy 2.4: 61 Kendall τ-b to 2e-16, ranking exact, Dirichlet mean + percentile-CI to 1e-12)
Analytic Hierarchy Process priority weights = normalised principal (Perron) eigenvector of a reciprocal pairwise-comparison matrix by power iteration, with the Saaty Consistency Index / Consistency Ratio and the CR<0.10 acceptance gate
Oracle
Saaty (1980) canonical Random Index table (RI(5)=1.12) reproduced exactly + SciPy/LAPACK scipy.linalg.eig (BSD-3-Clause) as an independent eigensolver computing the same uniquely-defined Perron eigenvector/eigenvalue, matched to <1e-9. The Pareto / sensitivity / MAUT decision layer built on this kernel stays MODELLED (the WSM/WPM/TOPSIS/VIKOR/PROMETHEE/ELECTRE aggregators are separately externally validated — see the rows below)
Tests
tests/mcda_ahp_reference.rs (Saaty 1980 Random Index table n=1..10 EXACT; priority vector + λ_max + CR vs SciPy/LAPACK eig on a consistent 3×3 and inconsistent 3×3 / 4×4, matched to <1e-9)
Weighted Sum Model (min–max-normalised additive aggregate + ranking) and Weighted Product Model (sum-normalised, reciprocal-for-cost multiplicative aggregate) — the two value-aggregation trade-study scorers
Oracle
pymcdm (methods.WSM + normalizations.minmax_normalization; methods.WPM + normalizations.sum_normalization) — an independent, widely-used third-party Python MCDA library computing the same uniquely-defined weighted aggregates; matched to <1e-9. Regenerable offline via tests/fixtures/mcda_wsm/ and tests/fixtures/mcda_wpm/. Garbage-in-garbage-out on the inputs; the sensitivity/robustness layer stays MODELLED
Tests
tests/mcda_wsm_reference.rs (WSM scores + ranking) and tests/mcda_wpm_reference.rs (WPM scores + ranking), each vs pymcdm to <1e-9 on a fixed 4×3 benefit/cost decision matrix
Technique for Order of Preference by Similarity to Ideal Solution — min–max normalisation, weighted positive/negative ideal solutions, relative closeness Cᵢ = d⁻/(d⁺+d⁻) and ranking
Oracle
pymcdm methods.TOPSIS + normalizations.minmax_normalization — an independent third-party MCDA library computing the same uniquely-defined closeness coefficients; matched to <1e-9. Regenerable offline via tests/fixtures/mcda_topsis/generate_topsis_reference.py
Tests
tests/mcda_topsis_reference.rs (closeness coefficients + ranking vs pymcdm to <1e-9 on a fixed 4×3 benefit/cost matrix)
VlseKriterijumska Optimizacija — group-utility Sᵢ, individual-regret Rᵢ and the compromise index Qᵢ at strategy weight v=0.5, with the lower-is-better ranking
Oracle
pymcdm methods.VIKOR(v=0.5) — an independent third-party MCDA library computing the same uniquely-defined range-normalised S/R/Q aggregation; matched to <1e-9. Regenerable offline via tests/fixtures/mcda_vikor/generate_vikor_reference.py
Tests
tests/mcda_vikor_reference.rs (Q index + ranking vs pymcdm to <1e-9 on a fixed 4×3 benefit/cost matrix)
Preference Ranking Organization METHod — pairwise preference index with the six standard generalised-criterion shapes, positive/negative outranking flows and the complete net-flow ranking (usual criterion validated)
Oracle
pymcdm methods.PROMETHEE_II('usual') — an independent third-party MCDA library computing the same uniquely-defined net outranking flow; matched to <1e-9. Regenerable offline via tests/fixtures/mcda_promethee/generate_promethee_reference.py. The thresholded (q/p/σ) preference shapes reduce to the same generalised-criterion algebra and stay property-checked
Tests
tests/mcda_promethee_reference.rs (net outranking flow + ranking, usual criterion, vs pymcdm to <1e-9 on a fixed 4×3 benefit/cost matrix)
ELimination Et Choix Traduisant la REalité — concordance / discordance matrices, the concordance-and-non-veto dominance relation and the outranking choice kernel (all-benefit, sum-normalised-weight, single-global-scale convention)
Oracle
pyDecision algorithm.electre_i (Valdecy Pereira) — an independent third-party multi-criteria decision library computing the same uniquely-defined concordance/discordance/dominance/kernel; matched element-for-element to <1e-9. Regenerable offline via tests/fixtures/mcda_electre/generate_electre_reference.py. The ĉ/d̂ threshold choices are analyst inputs (sensitivity stays MODELLED)
Tests
tests/mcda_electre_reference.rs (concordance, discordance, dominance matrices element-for-element + kernel/dominated set vs pyDecision to <1e-9 on a fixed 4×3 all-benefit dataset)
Weighted Aggregated Sum Product ASSessment (linear-normalised convex blend λ·WSM+(1−λ)·WPM at λ=0.5) and the MOORA ratio system (vector-normalised weighted benefit total minus weighted cost total) — the stability-hardened value blend and the signed ratio-system scorer
Oracle
pymcdm (methods.WASPAS + normalizations.linear_normalization, l=0.5; methods.MOORA ratio system, vector normalisation) — an independent, widely-used third-party Python MCDA library computing the same uniquely-defined aggregates; matched to <1e-9. Regenerable offline via tests/fixtures/mcda_waspas/ and tests/fixtures/mcda_moora/. Garbage-in-garbage-out on the inputs; the sensitivity/robustness layer stays MODELLED
Tests
tests/mcda_waspas_reference.rs (WASPAS preferences + ranking) and tests/mcda_moora_reference.rs (MOORA scores + ranking), each vs pymcdm to <1e-9 on a fixed 4×3 benefit/cost decision matrix
COmplex PRoportional ASsessment — column-sum-normalised benefit significance S⁺ plus the inverse-cost significance term, the relative significance Q and utility degree U = Q/max Q (best alternative = 1)
Oracle
pyDecision algorithm.copras_method (Valdecy Pereira) — an independent third-party MCDA library computing the same uniquely-defined COPRAS relative-significance/utility. pyDecision is used deliberately: pymcdm 1.4.0's COPRAS collapses algebraically to the trivial S⁺+S⁻ and is not a faithful reference, whereas pyDecision implements the canonical Q = S⁺+(min(S⁻)·ΣS⁻)/(S⁻·Σ(min(S⁻)/S⁻)). Matched to <1e-9; regenerable offline via tests/fixtures/mcda_copras/generate_copras_reference.py
Tests
tests/mcda_copras_reference.rs (utility degrees + ranking vs pyDecision to <1e-9 on a fixed 4×3 benefit/cost matrix)
Tabular-CUSUM detector worst-case detection latency (⌊h/(z−k)⌋+1) and out-of-control average run length, used by the timing-protection-level and spoof monitors
Oracle
Published tabular-CUSUM ARL: Siegmund (1985) Brownian-motion approximation (Hawkins & Olwell 1998, eq. 3.7) cross-anchored to Montgomery, Introduction to Statistical Quality Control (Wiley) ARL tables for k=½, h∈{4,5}; deterministic latency matched exactly to a first-passage oracle. The composed TPL bound stays MODELLED
Tests
tests/timing_protection_level_under_spoofing_reference.rs (16 deterministic-latency cases EXACT vs a first-passage oracle; 8 ARL₁ cases, Monte-Carlo @60k trials vs the Siegmund approximation + published Montgomery…
Coast phase-error variance growth q_wf·t + q_rw·t³/3 + q_drift·t⁵/20 and its monotone inversion to a timing-error holdover threshold
Oracle
scipy 1.18 (BSD-3-Clause): linalg.expm computing the Van-Loan 1978 discrete process-noise Q₀₀ = ∫₀ᵗ ΦQcΦᵀds via Padé scaling-and-squaring on the 6×6 augmented matrix — an independent route that never sees kshana's polynomial coefficients; plus optimize.brentq inverting the same monotone curve vs kshana's bisection. The per-class red-noise floor figures (ClockClass) stay MODELLED
Tests
tests/gnss_denied_clock_holdover_reference.rs (27 cases vs scipy 1.18: 12 coast-variance vs linalg.expm Van-Loan Q₀₀ worst rel 1.1e-15; 7 holdover inversions vs optimize.brentq worst rel 1.5e-16)
Effective architectural diversity = inverse-Simpson / Hill-order-2 number 1/Σpᵢ² over per-independence-group source qualities (the diversity term in the resilience score)
Oracle
scikit-bio 0.7.3 skbio.diversity.alpha.inv_simpson (McDonald et al., BSD-3-Clause) — an independent third-party library computing the uniquely-defined inverse-Simpson index; reproduced byte-for-byte. The DHS-RPCF scoring framework (weights/levels) built on top stays MODELLED
ValidatedAzimuth-dependent skyline mask: visible set, azimuth and elevation, and DOP under the maskskyline (Skyline::cut_deg,Skyline::visible); frames (look_angles,is_visible_skyline); orbit (UserPath,sky_view,cofactor_from_los); geometry (sky_geometry)
Capability
skyline + frames::look_angles + geometry::sky_geometry: the geodetic (WGS-84) azimuth and elevation of a satellite from a ground site, an azimuth/elevation mask table linearly interpolated in azimuth with wrap at 360 deg, the visible set el >= max(mask_deg, mask(az)), and the common-clock DOP and ENU-and-clock cofactor matrix of that set (the kernel behind orbit::dop). It is how the [skyline] section decides visibility in the gnss-sim, jamming, leo-pvt, leo-pass and nmea-scenario kinds and, with the ground-site user_site, in orbit and integrity
Oracle
Orekit 12.2 (Apache-2.0): a TopocentricFrame on a WGS-84 OneAxisEllipsoid for the look angles and an ElevationMask per table, run by xval/orekit-skyline/OrekitSkyline.java on committed Earth-fixed states (the propagation is not compared), with gnss_lib_py 1.0.4 (wheel metadata MIT) get_dop for the DOP and the ENU-and-clock DOP matrix of the sets Orekit finds visible. Pre-registered (xval/orekit-skyline/PREREGISTRATION.md, commit d94130d4) before either oracle was run on the comparison inputs; Orekit's linear-with-wrap interpolation was confirmed beforehand on a synthetic probe. Scope: table masks, geodetic look angles, the masked visible set and its common-clock DOP and cofactor. NOT checked here: the canyon, sector and terrain reductions and the presets (their 360-point tables entered as inputs; rows 'Street-canyon, sector and terrain reductions to a table' and 'Skyline presets'), the per-system clock states and the error ellipse (row 'Horizontal error ellipse and per-system cofactor from the geometry covariance'), the GMST rotation of the user_site ground site in orbit/integrity and its dating of each propagator to one instant (Orekit is given Earth-fixed states), and any real obstruction (line of sight only)
Tests
tests/skyline_orekit_oracle.rs (skyline_masks_agree_with_orekit_and_gnss_lib_py_as_pre_registered: 1173 look angles at six sites and four epochs, six mask tables, 142 visible sets; worst |d el| 1.3e-13 deg and |d az|…
ModelledStreet-canyon, sector and terrain reductions to a tableskyline (Canyon::cut_deg,Sector::contains,TerrainPoint::elevation_deg,reduce_terrain,SkylineCfg::build)
Capability
skyline: the street canyon in closed form (cut atan((h - a) / d) with d = (w/2) / abs(sin(az - s)), open ends for a finite length), sharp-edged sectors clockwise with wrap, and terrain points reduced to a table with the curvature drop d^2/(2 * 4/3 R); the skyline is their pointwise maximum with any table
Oracle
Internal: the closed forms are checked against hand-computed values (wall-top elevations of a stated street, the open-end condition, the curvature drop of a stated point). No external oracle: the design note scopes these reductions as MODELLED, and Orekit's ElevationMask takes a table, so the comparison of row 'Azimuth-dependent skyline mask: visible set, azimuth and elevation, and DOP under the mask' reads their 360-point tables as inputs and does not check them
skyline::PRESETS: four REPRESENTATIVE geometries with the design note's stated dimensions and a provenance line each: city-street (a canyon, street 0 deg, 20 m between 30 m walls, antenna 1.5 m), harbour-approach (sectors: open to seaward over 180 deg, quay buildings 8 deg over 90 deg, the breakwater 3 deg), fjord (terrain: walls 600 m above the antenna at 1.5 km either side of a north-south channel, curvature drop applied), container-terminal (sectors: stacks 12 deg over the landward half, crane booms 35 deg and 25 deg over two narrow sectors); overridable field by field
Oracle
None: chosen and stated dimensions, not a survey of any street, port, fjord or terminal (the aspect ratio of the street is inside the 0.5-2 band of Oke (1988); the fjord is of the order of the west-Norwegian fjords; the stacks use the ISO 668 high-cube height). The test checks the presets against their own stated dimensions, a transcription check, which the validation rule does not count
ValidatedHorizontal error ellipse and per-system cofactor from the geometry covariancegeometry (sky_geometry,ellipse); orbit (cofactor_from_los)
Capability
geometry::sky_geometry: the ENU-and-clock cofactor (H^T H)^-1 with one receiver clock in common or one per system (the orbit::cofactor_from_los kernel), and the horizontal error ellipse from its east-north block (semi-axes k sigma sqrt(lambda), k = sqrt(-2 ln(1 - p)), the major axis's azimuth from north)
Oracle
numpy 2.5.4 (BSD-3-Clause), xval/numpy-ellipse/ellipse.py: numpy.linalg.inv of G^T G and numpy.linalg.eigh of its horizontal block on the committed inputs, an independent numerical library performing the inverse and the eigen-decomposition by its own algorithms (standard P2). Pre-registered in xval/numpy-ellipse/PREREGISTRATION.md (commit 2dfc9c16) before the run; the driver's output formatting was corrected once after the pre-registration (a numpy scalar written as np.float64(...)), with no change to the protocol or the tolerances. The validated claim is the linear algebra on those inputs; the Gaussian error model and the UERE are stated inputs, not a measured accuracy or an integrity bound
Tests
tests/ellipse_numpy_oracle.rs (ellipse_and_cofactor_agree_with_numpy_as_pre_registered: 270 cases, the 142 committed visible sets of the skyline comparison under one common clock and under one clock per system; worst…
ModelledSky geometry functions: visibility, DOP, covariance, error ellipse, Monte Carlo cloudgeometry; orbit (cofactor_from_los,dop_cov,dop); surface (sky_geometry,sky_monte_carlo,skyline_presets,skyline_profile,skyplot_svg)
Capability
geometry (sky_geometry, sky_monte_carlo, skyline_presets, skyline_profile, skyplot_svg; CLI kshana sky and kshana skyline, Python and WASM functions of the same names, MCP sky_geometry): from a site, a time and constellation presets (or satellites as azimuth/elevation or ECEF), the satellites visible and masked under a skyline, GDOP/PDOP/HDOP/VDOP/TDOP, the unit-variance cofactor (H^T H)^-1 and the covariance in east-north-up with one receiver clock per system or one in common, the horizontal error ellipse (semi-axes k sigma sqrt(lambda), k = sqrt(-2 ln(1 - p)), orientation 0.5 atan2(2 Q_EN, Q_NN - Q_EE) from north), and n seeded least-squares errors under independent Gaussian range errors with their empirical covariance, RMS and 95 % containment. The cofactor is orbit::cofactor_from_los, the one kernel behind orbit::dop and orbit::dop_cov (dop is dop_cov's reduction): no second DOP implementation
Oracle
The single-clock DOP factors rest on the kernel validated against gnss_lib_py 1.0.4 (row 'GNSS geometry / dilution of precision (DOP)'); this row's own checks are internal: sky_geometry reproduces orbit::dop and dop_cov on the same geometry, the ellipse reproduces an analytic 2x2 eigen-decomposition, ECEF and azimuth/elevation inputs agree, and the seeded cloud's RMS and 95 % containment agree with the geometry. The ENU covariance with one clock per system and the ellipse are VALIDATED in row 'Horizontal error ellipse and per-system cofactor from the geometry covariance' (numpy, P2); what stays MODELLED here is the composition: the preset propagation to a representative sky, the Monte Carlo cloud (checked against the geometry it samples, a self-consistency check) and the surfaces. The constellation presets give a representative sky, not a dated one; the Gaussian error model and the UERE are stated inputs, not a measured accuracy or an integrity bound
clock_library: a device card (white phase, white, flicker and random-walk FM plus a linear frequency drift) fitted to one third of a named measured clock record after a logged conditioning pass (gaps, phase outliers, phase steps, bursts removed; frequency steps logged), weighted by the lag-1-identified noise-type degrees of freedom, and its predicted Allan deviation scored on the held-out two thirds; readers for RINEX and IGS clock files, BIPM per-laboratory files and Circular T Section 1; conversion of a card to the slot-timing noise model, the extended Kalman clock model and the spoofing monitor's noise levels
Oracle
Measured records, held out: IGS final combined 30 s clocks of the 11 GPS Block IIF satellites over 2026-03-01 to 14 (a window no test had opened; IGS, open with attribution) and the JammerTest 2024 u-blox ZED-F9P receiver in its non-scored stationary sessions (GPL-3.0-or-later). D9 external comparison (pre-registered 14169982, every card's predicted / measured Allan deviation within [1/1.5, 1.5] at its fitted averaging times), a finding (stays MODELLED): 10 of 11 GPS IIF cards are within the bar; G27 is optimistic at the two-hour scale (0.662 at 7680 s, worst factor 1.510), the periodic hour-scale error a power-law card does not carry. The receiver TCXO card is blocked: outside the logged transmissions the dataset's other sessions leave 167 epochs against the 3600 required. Reported, not blind (records opened by earlier rows): the 5071A caesium card is within the bar (worst 1.070 up to 4096 s), the OCXO card is conservative and outside it (2.348 at 128 s; its noise floor changes during the record), and the Norcia strontium card fitted on the three shortest published points is within it at the two scored points (1.291, 1.405). Galileo passive hydrogen masers (no login-free source of final Galileo clocks reachable) and the Deep Space Atomic Clock (one open in-space value) are blocked. D9 round 2, the ZED-F9P receiver TCXO as a model class on 14 days of 12 static Wroclaw stations (Zenodo 6488497, CC BY 4.0; pre-registered 698f3842), a finding: with the registered pipeline 11 of 12 stations were not evaluable (RAIM rejected most ionosphere-free epochs) and BX14 failed; a disclosed re-run with a corrected pipeline (pre-registered cdfde420) passes 1 of 11 blind stations (BX12, 1.412; BX07 1.532 and BX13 1.542 just outside), because single-point receiver clocks from intermittent 30 s files are not a clean oscillator record
clock_library::card::DeviceCard::fit_phase_periodic: a device card that carries the once- to four-per-revolution phase terms (period T/k, T = 43 082.05 s), fitted jointly with the quadratic by least squares on one third of a measured satellite clock record, with each term's closed-form Allan contribution 4 A^2 sin^4(pi tau / P) / tau^2 in the predicted Allan deviation
Oracle
Measured records, held out: IGS final combined 30 s clocks of the 11 GPS Block IIF satellites over 2026-04-01 to 14 (a window no test had opened; IGS, open with attribution). D9 external comparison (pre-registered 698f3842, predicted / measured Allan deviation within [1/1.5, 1.5] at every fitted averaging time on every satellite), a finding (stays MODELLED): 9 of 11 cards are within the bar; G03 (1.947) and G25 (1.581) are optimistic between 60 s and 2000 s because their held-out records are noisier than their fit thirds (G25's held-out part carries 16 gaps and 29 phase outliers), a change of the clock rather than a missing periodic term. Dropping the periodic terms from the prediction fails all 11 cards (1.77 to 4.96), so the terms carry the prediction
filterpy 1.4.5 KalmanFilter (R. Labbe, MIT), with F via scipy.linalg.expm and Q via the Van-Loan 1978 block-matrix — an independent reference implementation reproducing kshana's full filter trajectory. Cross-implementation consistency: the clock physics / Allan calibration are not externally validated, so this stays MODELLED. 0.30 external comparison, a finding (stays MODELLED): on IGS final 30 s clocks of 11 GPS Block IIF satellites (2025-08-17 to 30), with Q fitted to the first-half ADEV and the filter scored on the held-out second half, the filter is consistent one step ahead (0.906 to 0.971 of epochs inside its 95 % band, bar 90 %) but over-confident at one hour on 9 of 11 clocks (0.486 to 0.829) and its innovation sums fall outside [2.4, 3.6] on 5. The fitted white-phase noise is zero on every satellite, and the hour-scale error (periodic terms, flicker FM) is not in the 3-state model. 0.30 round 2, an extended filter (flicker FM and once- and twice-per-revolution states) on fresh held-out IGS clocks of 2025-09-01 to 14 (pre-registered d8d18573), a finding (stays MODELLED): the one-step and one-hour consistency criteria hold on all 11 satellites (0.948 to 0.988 and 0.907 to 0.996), but the innovation-sum criterion fails on G24 (0.964) and G30 (1.929) against [2.4, 3.6]. D9 round 3, the round-2 extended filter, tuning and criteria unchanged after the frozen clock_library conditioning detector, on fresh IGS clocks of 2026-04-01 to 14 (pre-registered 698f3842), a finding (stays MODELLED): (a) one-step 0.941 to 0.968 and (c) triple-NIS mean 2.641 to 3.549 hold on all 11 satellites; (b) one hour fails on G09 (0.894) and G26 (0.876) against 0.90
Tests
clock_state::tests (analytic van-Loan Q; NEES; PSD positivity); tests/clock_state_reference.rs (full predict+update trajectory — state x and 3×3 covariance P over 1925 steps / 4 parameter sets vs filterpy 1.4.5; worst…
Sagnac magnitude checked against an authoritative PUBLISHED VALUE — N. Ashby, 'Relativity in the GPS', Living Reviews in Relativity 6:1 (2003), Eq. 1.29: an eastward equatorial circumnavigation accrues 207.4 ns (2ωA_E/c²); kshana reproduces 207.386 ns. Independently corroborated by RTKLIB 2.4.3 geodist() (Takasu, BSD-2-Clause), which carries the same 2Aω/c² geometry. The composite BIPM TWSTFT transponder/common-view/PPP budget has no external oracle, so the capability stays Modelled
Tests
timetransfer::tests (reciprocal cancellation; two-form Sagnac identity); tests/time_transfer_error_budgeting_reference.rs (equatorial-circumnavigation Sagnac = 207.386 ns vs the published Ashby 207.4 ns to <0.05 ns…
GPS C/A Gold cross/auto-correlation matched EXACTLY (integer ±65/−1/63) against independent IS-GPS-200 code generation; BPSK-R(1)/BOC(1,1) PSD shape vs an independent scipy periodogram. The modulation/SSC/DLL closed forms (Betz 2001 / Kaplan & Hegarty) remain analytic, so the row stays MODELLED — but the code-correlation sub-claim is externally matched. 0.30 round 2, against the printed values of Betz 2001 and Hein et al. 2006 (pre-registered 1181ce1d, half a printed unit), a finding (stays MODELLED): all 45 Table 1 spectral cells, the 1 GHz correlation values, the MBOC interference figures and 2 of 3 Fig. 17 offsets agree; the 24 MHz first side lobes, the BOC(8,4) Fig. 17 offset and every Figs. 18-21 worst-case multipath bias (11 to 15 % low) do not. Kshana and an independent NumPy evaluation of the pre-registered reading agree, so the gap lies in that reading or in an unstated detail of the paper; cause unresolved
Tests
navsignal::tests (BPSK self-SSC = 2/3R_c; unit-area PSD; DLL); tests/nav_signal_modulation_code_tracking_reference.rs (GPS C/A Gold cross/auto-correlation exact-integer match vs independent IS-GPS-200 code generation…
ValidatedGPS L1 C/A spreading-code generationsdr
Capability
G1/G2 LFSR C/A-code generator reproducing the IS-GPS-200 published first-10-chip octal verification vectors for PRN 1–9 (plus 1023-chip length, 512-ones balance, three-valued periodic autocorrelation)
Oracle
IS-GPS-200 Table 3-Ia 'Code Phase Assignments' — the authoritative US-Government public-domain published first-10-chip OCTAL verification vectors (GPS Directorate, navcen.uscg.gov); kshana's own G1/G2 LFSR regenerates every PRN 1–9 vector exactly. The downstream modulation/SSC/DLL closed forms stay Modelled (separate row)
Cold-atom interferometer accelerometer from first principles (k_eff·T², QPN)
Oracle
Published CAI primary-paper numeric vectors (Cheinet 2008 transfer function; Peters/Freier sensitivity): k_eff·T² matched exactly, shot-noise ASD a one-sided floor within ~2× of each published instrument (real devices carry technical noise above the quantum floor). A bracket, not parity. 0.30 round 2 (pre-registered e7de532a), a finding (stays MODELLED): the QPN-only rotation noise against the measured Gauguet et al. 2009 Fig. 14 points gives ratios 0.695 to 0.788 at the three admitted points (0.751 at the operating point), missing the pre-registered 30 % bar by 0.5 points; the QPN floor sits 21 to 31 % below the measured noise, as expected of a model without laser and detection noise. The gradiometer composition matches the Janvier et al. 2022 QPN model line (ratio 1.013 to 1.017, Reference). The measured test cannot discriminate a sqrt(2) composition error; only the unequal-contrast Janvier line can
Tests
quantum_imu::tests (k_eff; Mach-Zehnder T²; Freier-2016 floor bracket); tests/quantum_inertial_sensor_reference.rs (transfer function |H(ω)|, k_eff·T² and shot-noise ASD vs published Cheinet 2008 / Peters / Freier…
Mach–Zehnder fringe-ambiguity dynamic range: the 2π-periodic fringe readout sets a maximum unambiguous specific force a_max=π/(k_eff·T²), and the unambiguous range in resolution cells a_max/σ_a=π/σ_Φ is independent of the optical scale factor — the T² sensitivity gain costs unambiguous range in exact lockstep (interrogation time trades resolution for range, leaving the cell count fixed by the readout phase noise)
Oracle
Self-consistency of the interferometer fringe model: the half-fringe edge, the 2π-periodic aliasing structure, and the scale-factor cancellation in the range/resolution ratio are closed-form algebraic identities checked against the engine's own Mach–Zehnder phase and sensitivity functions — internal-consistency checks, NOT an external dataset, so the row stays InternalConsistency. MODELLED ideal three-pulse fringe-ambiguity; no wavefront-aberration or contrast-loss bounds on the unambiguous range
Tests
quantum_imu::tests (a_max sits at the ±π half-fringe edge with the 1/T² range scaling; wrapped-phase recovery is exact inside [−a_max,a_max] and aliases by exactly 2·a_max outside it; the unambiguous dynamic range…
Composed bias + scale-factor + VRW + stability-decay position budget over holdover
Oracle
Independent numpy Monte-Carlo SDE integration of double-integrated white-acceleration noise (validates the analytic VRW variance by a genuinely independent algorithm) + a Groves 2013 published-value anchor for the bias/scale-factor terms; the CAI device numbers quantify partner hardware and stay MODELLED
Tests
budget_tests (bias vs AccelModel integrator; VRW vs analytic integral); tests/quantum_inertial_dead_reckoning_reference.rs (VRW vs an independent numpy Monte-Carlo double-integration of white-acceleration noise, worst…
15-state error-state EKF, loosely coupled (validated on a real IMU/GNSS drive against NaveGo); a tightly coupled pseudorange/Doppler UKF and a coupled clock+position filter (consistency-only, outside the validated claim)
Oracle
NaveGo v1.4 (R. Gonzalez, LGPL-3.0, run as a tool under GNU Octave, never linked): its loosely coupled ins_gnss on its own real Ekinox IMU/GNSS land-vehicle drive, both solutions scored against the Ekinox reference trajectory. Both filters read the same 20 Hz float32 input, and Kshana's configuration is mapped mechanically from the dataset's published parameters. Kshana's horizontal and vertical position RMS and its position-innovation RMS are within 1.2x of NaveGo's (measured 0.912, 0.940 and 0.861) and its mean normalised innovation is at most 1.5 (measured 0.395: about 2.5x under-confident, given the dataset's stated GNSS sigmas); tolerances fixed before the comparison. Scope: GNSS is present at 5 Hz for the whole drive, so no outage or coast is tested and the check catches gross filter defects rather than fine tuning (lever-arm-corrected raw fixes would land near the vertical bar); the lever-arm transform is computed in the test harness. The tightly coupled UKF and the coupled clock+position filter keep their filterpy 1.4.5 consistency checks (linear posteriors to 2.4e-12) and stay outside the validated claim
Tests
fusion::tests (UKF==linear-KF identity; outage coast; NEES); tests/gnss_ins_sensor_fusion_reference.rs (50 cases vs filterpy 1.4.5: linear EKF loose/tight + coupled-PNT posteriors to ≤2.4e-12; UKF 40-epoch run worst…
ModelledGNSS-denied jamming resiliencejamming
Capability
Geometry J/S link budget, anti-jam C/N₀, per-satellite loss-of-lock
Oracle
Anti-jam C/N₀ link-budget equation cross-checked against an independent numpy re-derivation (shares the same closed form → InternalConsistency) plus a real-JammerTest-2024 C/N₀ degradation characterisation. 0.30 round 2, blocked (stays MODELLED): the measured JammerTest 2024 C/N0 drop against a link-budget prediction (pre-registered 6b2f6240, within 3 dB) was not run because the F8.1 antenna position and pointing at Bleik and the L1 share of the multi-band 1.6.4 ramp power are not documented in the test catalogue or the official log
Tests
jamming::tests (broadband_q_equals_the_spectral_separation_closed_form and cw_q_equals_the_spectral_separation_closed_form: the Q table, CW and narrowband 1, broadband and swept 2, pinned to the navsignal closed forms…
TEXBAT scenario parameters (Humphreys 2012) — characterisation, not pinned vectors. 0.30 external comparison, a finding (stays MODELLED): on JammerTest 2024 (Bleik; GPL-3.0 data) the observable-level monitors (clock-aided chi-square, RAIM, solve failure) raised no alarm within 10 s of any of the 8 evaluable published spoofing onsets (where the pre-onset window was clean, the first alarm came 10.2 to 39 s after the published slot start), and three pre-onset windows carried false alarms from the clock monitor (81, 64 and 173 epochs). The first observable effect was a loss of dual-frequency GPS tracking, not a RAIM inconsistency. The published onsets are minute-resolution schedule slot starts; that they precede the RF capture is an interpretation, not a measurement. 0.30 round 2, against the organisers official JammerTest 2024 log (second resolution; pre-registered 6b2f6240 with a Hadamard-calibrated three-state clock monitor without latching), a finding (stays MODELLED): the monitor alarms within 10 s at 4 of 10 logged onsets; at the other six it is 18 to 211 s late or raises pre-onset clock false alarms (1 at 2.1.4, 85 at 2.3.15). D9 round 3 (a receiver card from the JammerTest unit's other sessions, pre-registered 14169982) blocked: 167 training epochs against 3600. D9 rounds 4 and 4b (noise levels from a u-blox ZED-F9P model-class card, Wroclaw 2021, pre-registered 698f3842; 4b a disclosed re-run with a corrected extraction, cdfde420), a finding (stays MODELLED): zero pre-onset false alarms at all 10 logged onsets and 6 of 10 within 10 s; 2.1.1, 2.3.5, 2.3.10 and 2.6.1 stay late (+211, +23, +18, +51 s), unchanged by the clock noise level
Tests
tests/spoof_texbat_validation.rs (TEXBAT parameter characterisation); tests/spoof_detection_jammertest_oracle.rs::monitors_against_the_published_onsets_reproduce_the_recorded_disagreement (pins the finding)…
ModelledTiming Protection Level under spoofingtpl
Capability
Closed-form bound on worst-case undetected time error = monitor floor + oscillator coast-σ over CUSUM detection latency, reported as a red-noise-floor band
Oracle
Composes Validated primitives (allan/holdover van-Loan, security floor); calibrated on JammerTest 2024 scenario 2.1.1 (~1.01 ms real served-time pull vs ≤51 ns claimed). Bridge over Validated parts — not itself an external validation. 0.30 external comparison, a finding (stays MODELLED): on the JammerTest 2024 spoofing scenarios the measured undetected served-time error exceeds the nominal TPL at 3 of 8 detected onsets: 184.3 vs 34.6 ns (2.1.4), 295.7 vs 159.3 ns (2.3.5) and 143.0 vs 69.0 ns (2.3.11); three of the five inside are trivial (a latched pre-onset false alarm). Likely cause: the coast term from a 60 s calibration with zero drift noise does not cover the receiver TCXO's frequency wander over 15 to 39 s. The measured error is Kshana's own single-point clock estimate against a line fitted to its own pre-onset estimates, so even agreement would not have been fully independent. 0.30 round 2, against the official JammerTest 2024 log onsets and the measured receiver time error (pre-registered 6b2f6240), a finding (stays MODELLED): at 4 of 10 logged onsets the undetected error exceeds the drift-state TPL (634.7 against 458.5 ns at 2.1.1, 232.6 against 16.2 ns at 2.3.5, 41.3 against 6.0 ns at 2.3.10, 218.6 against 47.4 ns at 2.6.3)
NASA/JPL Three-Body Periodic Orbit Database (SSD, periodic_orbits.api; Earth–Moon L2 Southern halo family, Howell/Davis methodology) — externally-published period T, Jacobi C and perpendicular-crossing initial conditions; kshana's single-shooting STM corrector, seeded with the catalog IC and perturbed, converges back onto the catalog members
Tests
cr3bp::tests (STM vs finite-diff); tests/cislunar_mission_analysis_reference.rs (5 JPL L2-S NRHO members: the 9:2 + 4 neighbours; worst |ΔC| 1.5e-5, |ΔT|/T 9.6e-5, perilune 0.7 km in JPL length units)
Gravity-map matching, terrain-referenced (TERCOM/SITAN), magnetic anomaly
Oracle
IGRF-14 coefficients; first-principles matched-filter CRLB. The real-field gravity benchmark's field is externally checked (row 'Gravity-anomaly synthesis from a spherical-harmonic model'), but its navigation outcome has no external oracle
Tests
tests/alternative_complementary_pnt_reference.rs (map-matching CRLB; IGRF-14 field); mapmatch::tests; gravimeter::tests; igrf::tests; tests/gravity_map_real_field.rs (the 60-min GPS-denied benchmark on the real EGM2008…
ValidatedGravity-anomaly synthesis from a spherical-harmonic modelgravity_sh,altpnt::maps,gravimeter
Capability
Gravity anomaly in spherical approximation (Barthelmes STR09/02 eq. 126, ICGEM gravity_anomaly_sa) from fully normalised coefficients against a closed-form level ellipsoid (WGS84/GRS80), with the zero-degree term, gridded for the gravity-map matcher
Oracle
ICGEM Calculation Service (GFZ Potsdam), an independent implementation: EGM2008 truncated at degree 70, gravity_anomaly_sa, WGS84, tide-free, zero-degree term, h = 0. The service output carries no explicit licence, so the grids are fetched at test time with a pinned SHA-256 (tools/fetch_icgem_grid.sh) and only ten values are cited with attribution
Tests
tests/icgem_gravity_anomaly_oracle.rs (egm2008_n70_matches_cited_icgem_values in CI on ten cited ICGEM nodes; egm2008_n70_gravity_anomaly_sa_matches_icgem over the full fetched 0.25-deg grids of Chad 9-15N 19-25E and…
ValidatedSRTM digital-elevation reader on real terrainaltpnt::terrain
Capability
Hand-rolled SRTM .hgt parser (16-bit big-endian, north-row-first, void-aware) + bilinear sampler reading a real public-domain DEM tile and resolving a documented survey benchmark
Oracle
NASA/USGS SRTM v3 (1-arc-second) N36W117 tile — US-Government PUBLIC-DOMAIN elevation data from the AWS elevation-tiles-prod open mirror, decimated to 6-arc-second and committed under tests/fixtures/terrain/ (see NOTICE.md), plus full-resolution crops of N36W117 and N36W119 cross-checked node-for-node against GDAL's independent SRTMHGT reader and the Mount Whitney (4421 m) and Badwater (−86 m) survey benchmarks. The documented Badwater Basin benchmark (−86 m, lowest in North America) anchors the geo-referenced read. The terrain-matching/TERCOM nav-fix that consumes the DEM stays Modelled (separate Alternative/complementary PNT row)
Tests
tests/terrain_nav_validation.rs (real_srtm_committed_badwater_tile_reads_real_relief — the committed 6-arc-second decimation of the public-domain SRTM v3 N36W117 tile places Badwater Basin, the lowest point in North…
ModelledReproducibility & software assurancereport,scenario; CI (golden/determinism/SBOM)
Capability
Deterministic, scenario-hashed, SBOM + cross-platform golden gates
Oracle
SBOM conformance to the official CycloneDX 1.5 JSON Schema (+ valid SPDX identifiers) — an external published standard, zero validation errors over the full dependency graph; the FoM-determinism / byte-reproducibility part remains a pinned self-consistency check, so the row stays MODELLED
Tests
tests/golden.rs, tests/determinism.rs, tests/cross_platform_golden.rs; tests/reproducibility_software_assurance_reference.rs (the generated SBOM validates with zero errors against the official CycloneDX 1.5 JSON Schema…
Labelled synthetic impairment corpus + detector-agnostic ROC/AUC/confusion/Pfa-Pmd harness; leakage guard, stratified split, distribution-shift (in- vs out-of-regime) optimism report. Runnable from the CLI/bindings as the impairment-eval scenario kind (scenarios/impairment-eval.toml)
Oracle
Closed-form AUC bounds (Mann–Whitney) + a perfect-oracle detector; corpus is SYNTHETIC (parameter-grounded, not field/IQ)
ModelledAI/ML RF-impairment optimism-gap study & ID-only gap predictorimpairment_study,impairment_ml,eval_stats
Capability
Controlled synthetic study of the in-distribution→out-of-distribution AUC optimism gap across published-method and learned (logistic-regression / one-hidden-layer MLP) detectors: per-class scaling-law trends (Spearman ρ + slope on 1−severity) and an ID-only ridge predictor that estimates the gap from in-distribution diagnostics alone, scored leave-one-detector-out and leave-one-class-out. Reproducible via cargo run --release --example optimism_study
Oracle
Hand-derived statistics vs closed forms (binormal AUC Φ(d'/√2), DeLong variance, tied-rank Spearman, exact OLS recovery) + leave-one-out CV against the predict-the-mean baseline. Corpus is SYNTHETIC (parameter-grounded, never field/IQ) and the optimism gap is a synthetic→synthetic severity shift, NOT a sim-to-field result
Tests
impairment_study::tests (per-class oracle AUC≈1, learned optimism gap>0, grid shape + bootstrap CI brackets the mean + positive scaling trend, ID features finite, gap predictor beats predict-the-mean under BOTH…
Measured-ADEV ingestion (NNLS), trade table (timing/inertial holdover + benefit), resilience-vs-time envelope; floor caveat carried on the artifact. Runnable from the CLI/bindings as the quantum-trade scenario kind (scenarios/quantum-trade.toml)
Oracle
The measured-ADEV→PSD fit (NNLS) kernel is matched to scipy.optimize.nnls (tests/scipy_reference.rs / tests/quantum_vs_classical_pnt_trade_reference.rs) — an independent external kernel; but the trade NUMBERS quantify (never validate) a partner clock/CAI, so the trade itself stays MODELLED, no validation halo
Tests
quantum_trade::tests (ADEV round-trip recovery, NNLS non-negativity, floor-caveat present/absent, benefit>1, monotone envelope + alt-PNT bound); dominance_demonstrators (measured-ADEV is data-driven not floor-assumed…
Solar/geomagnetic indices (definitional Kp↔ap table), the Jacchia-1971 exospheric temperature, and the Jacchia 1971 thermospheric density (SAO Special Report 332: static diffusion profiles reproducing its Table 7, with the diurnal, geomagnetic, semiannual, seasonal-latitudinal and helium variations) against the static USSA76 atmosphere. Runnable from the CLI/bindings as the space-weather scenario kind (scenarios/space-weather.toml)
Oracle
Definitional Kp↔ap table + Jacchia-1971 exospheric-temperature closed form (matched to <1 K vs the published anchors, tests/space_weather_reference.rs); the density is characterised against pymsis NRLMSISE-00 (an independent NRL model): the 800 km solar-cycle swing is 10.4x against NRLMSISE-00 13.6x. 0.30 revision: the calibrated first-order activity factor was replaced by Jacchia 1971 (default scenario density at 400 km 4.176e-12 -> 6.929e-12 kg/m3; reference_exospheric_temperature_k removed, mean_exospheric_temperature_k added). 0.30 round 2, measured TOLEOS CHAMP/GRACE/GRACE-FO accelerometer densities (ESA Swarm DISC; pre-registered bb8756e3), a finding (stays MODELLED): orbit averages at 400 to 500 km are within a factor 2 at solar maximum outside storms (CHAMP 2001 521/521, GRACE-A 2002 120/120, GRACE-FO 2024 526/547), but Jacchia 1971 is about 1.7x too dense in the 2008 solar minimum (GRACE-A 412/544 within a factor 2), so the density is not validated
Tests
space_weather::tests (Kp↔ap exact at grid points + round-trip + monotone, daily-Ap mean, exospheric-T vs published solar-min/mean/max + storm increment anchors, density unity-at-reference, solar-cycle swing in the…
CCSDS 502.0 OEM importer (parse_oem), tolerant of COMMENT lines / extra metadata keywords / covariance blocks and the exact inverse of the writer; round-trip + external-file ingest with a velocity-consistency check. Runnable from the CLI/bindings as the oem-interop scenario kind (scenarios/oem-interop.toml)
Oracle
Independent third-party CCSDS-502 parser oem 0.4.5 (MIT, https://pypi.org/project/oem/) — a separate codebase that decodes kshana's emitted EME2000/UTC OEM byte-identically (24 states across 2 fixtures, pos/vel Δ = 0) and whose strict reader confirms the metadata tokens; kshana's parser likewise agrees with it on a vendored external OEM. Two honest interop findings (the oem library rejects kshana's per-satellite multi-segment convention and its multi-entry covariance lines) are documented in the test — so this validates the conformant single-object interchange, not full CCSDS-502 conformance of every kshana variant
Tests
tests/ccsds_oem_interop_reference.rs (24 states / 2 fixtures decoded byte-identically by the independent oem parser, pos/vel Δ = 0); oem::tests (parse an external-tool OEM with extra keywords/comments/covariance…
Two-body launch azimuth(s) (sin Az = cos i / cos lat), minimum reachable inclination, circular velocity, Earth-rotation eastward bonus, dogleg plane-change Δv and daily opportunities, and, given an epoch and IERS finals2000A rows, the site speed about the true pole (launch::site_rotation_speed_at). Runnable from the CLI/bindings as the launch-window scenario kind (scenarios/launch-window.toml)
Oracle
Closed-form spherical-trig launch geometry vs published worked-example anchors (Vallado, Fundamentals of Astrodynamics 4th ed., Algorithm 37 launch-azimuth + Ch.6 plane-change; tests/launch_window_ascent_geometry_reference.rs). These re-use the same closed form kshana implements (a published-value parity / transcription check, InternalConsistency); MODELLED two-body, no rotating-Earth velocity-triangle / ascent / drag-loss model. 0.30 external comparison against Orekit 12.2 (Apache-2.0), a finding (stays MODELLED): the geometry agrees to rounding (inclination read back from the burnout azimuth over 158 cases, worst 2.6e-15 rad; minimum inclination, circular velocity and dogleg delta-v to 3e-16), but the Earth-rotation site speed misses the pre-registered 1e-6 relative at 62.9 deg latitude (1.12e-6). The gap grows as tan(lat), consistent with a polar-motion pole offset of about 5.8e-7 rad in Orekit's Earth frame that the spherical ω·R·cos(lat) omits. The claim was not narrowed before the run, so the row cannot promote on this comparison. 0.30 round 2 (pre-registered 60d7f773): with an epoch and IERS rows, the true-pole site speed (scenario field site_rotation_speed_true_pole_m_s) agrees with Orekit within 1e-6 at 24 sites over four longitudes (worst 4.9e-8), but the nominal no-epoch speed the scenario emits by default as the Earth-rotation bonus still misses by 1.12e-6, so the row stays MODELLED pending a decision on the default output
Peak deceleration (ballistic-coefficient-independent), velocity + altitude at peak-g, and peak-heating velocity for an exponential-atmosphere ballistic entry. Runnable from the CLI/bindings as the reentry scenario kind (scenarios/reentry.toml)
Oracle
Closed-form Allen–Eggers analytic entry, additionally cross-checked vs a scipy 1.18 solve_ivp (DOP853) numerical integration of the SAME drag-only entry ODE (tests/ballistic_re_entry_corridor_reference.rs, 36 cases, worst a_max rel 2.9e-9) — a numeric-integral-vs-own-analytic-form check, so still InternalConsistency, NOT an external validation. MODELLED ballistic (no lift), no aerothermal/TPS — heating output is a velocity, not a heat-flux. 0.30 external comparison, a finding (stays MODELLED): against the Stardust SRC entry reconstruction (Desai and Qualls, NASA NTRS 20080008567), the Allen–Eggers peak deceleration for the printed inertial entry state (12.9 km/s, -8.2 deg, H 7200 m) is 61.8 g against the best-estimated-trajectory 32.89 g, +88 % against a 15 % tolerance: the constant-flight-path-angle, no-gravity solution roughly doubles the peak of a shallow, faster-than-escape entry. Stardust carried no accelerometer, so the reconstruction is a trajectory estimate. 0.30 round 2, blocked (stays MODELLED): an integrated planar point-mass solution (reentry::simulate_planar_entry, US76, inverse-square gravity, curvature) agrees with the Stardust and Genesis best-estimated trajectories within 15 % (+11.3 %, +5.6 %), but those maxima are agency simulation outputs, not measurements, so this is a cross-check only. The accelerometer comparison (pre-registered 18863477) is blocked on a readable copy of the Hayabusa2 REMM analysis (doi 10.57350/jesa.16)
Tests
reentry::tests (peak-g independent of ballistic coefficient + physical g-band, grows with steeper γ / faster entry, peak-g velocity = V_e·e^(−1/2) and peak-heating = V_e·e^(−1/6) faster, peak-g altitude physical +…
SMAD space-triangle geometry: Earth angular radius, swath width, nadir GSD, maximum off-nadir access and circular period, plus the first-order J2 nodal period (eo_payload::j2_nodal_period) and the equatorial ground-track spacing R_e (omega_E - dOmega/dt) T_nodal (eo_payload::ground_track_spacing_equator_j2) with a contiguous-coverage flag, for an orbit altitude, inclination and sensor FOV/IFOV. Runnable from the CLI/bindings as the eo-coverage scenario kind (scenarios/eo-coverage.toml); its default inclination is the sun-synchronous one from eo_payload::sun_synchronous_inclination, a convenience default that is not part of the validated claim
Oracle
Orekit 12.2 (Apache-2.0) and GeographicLib 2.1 (MIT), both run as separate programs, pre-registered (80513191) before the engine change and any oracle run: sphere quantities within 1e-9 relative (worst: angular radius 5.4e-16, swath 7.0e-13, access 1.9e-14), period 1e-12 (1.6e-16), nadir GSD 1e-7 (2.9e-10), the J2 nodal period and node spacing within 0.1 % of an Orekit numerical J2 propagation with a node detector (1.2e-5 and 6.7e-5), the contiguous flag exact (168/168), and against the WGS-84 ellipsoid the limb within 0.3 deg (0.186 deg) and the maximum ground range within 0.5 % (3.2e-3). Removing the nodal regression term fails 35 of 42 node points. MODELLED scope: spherical-Earth footprint; no radiometry, MTF, atmosphere, jitter or glint; the sun-synchronous default inclination is not compared. The earlier Skyfield/SGP4 cross-check is kept as a regression. 0.30 revision: the spacing now carries the J2 nodal regression (default scenario 2756.37 -> 2752.17 km)
Tests
eo_payload::tests (angular radius 64° at 700 km + shrinks with altitude, nadir→zenith/zero-range, horizon→ε=0/max central angle, past-horizon errors, swath grows with FOV / GSD with altitude, ~2750 km node spacing…
ValidatedCCSDS Space Packet (133.0) TM/TC framingspace_packet
Capability
CCSDS 133.0-B Space Packet primary-header encode/decode (version/type/sec-hdr/APID/seq-flags/count/data-length) + a framing scenario. Runnable from the CLI/bindings as the space-packet scenario kind (scenarios/space-packet.toml)
Oracle
spacepackets 0.32.0 (us-irs/spacepackets-py, R. Mueller, Apache-2.0) — an independent third-party implementation of CCSDS 133.0-B-2; kshana's encode_packet/decode_packet matched byte-exact (the 6-octet primary header for 33 cases + the full encoded packet for 12)
Tests
space_packet::tests (header bits match the CCSDS-133 layout, encode→decode round-trips all fields, out-of-range/truncated rejected); tests/ccsds_space_packet_reference.rs (33 cases vs spacepackets 0.32.0, incl. 12…
ValidatedCCSDS Space Packet secondary header, packet error control and segmentationspace_packet (TimeCode,SecondaryHeader,packet error control,segmentation and reassembly,stream decoding)
Capability
CCSDS 133.0-B-2 Space Packet extensions beside the primary header: a secondary header carrying a CCSDS 301.0-B time code (CUC or CDS, explicit or implicit P-field) and ancillary data, the CRC-16 packet error control field, segmentation of a large user-data unit and its reassembly, stream decoding with sequence-gap and order checks, and PUS-TM packets. The primary-header output is byte-identical to before
Oracle
Pre-registered (1bd991e9) before any comparison; four independent implementations, none GPL, nothing vendored: ccsdspy 2.0.1 (BSD-3-Clause) decodes 34 Kshana-written packets in 5 streams (primary header, P-field, coarse and fine time, CDS fields, ancillary data, user data, PEC) and both 1000-octet segmented units are rebuilt from its user data with 0 mismatches, and it flags the dropped and the swapped packet streams that Kshana rejects; the CRC of 8 messages and 57 PEC fields equals CPython binascii.crc_hqx and spacepackets-rs 0.18.0 (check value 0x29B1); 66 CUC and CDS time-code cases equal spacepackets-rs 0.18.0 write and read octet for octet (3 CDS cases also spacepackets-py 0.32.0); 4 spacepackets-py 0.32.0 PusTm packets are byte-identical to Kshana's. Not covered: agency-epoch time codes (7 cases rest on unit tests), TC transfer frames and any link layer below the packet
Tests
tests/space_packet_ext_oracle.rs (five pre-registered bars: CRC, time codes, streams, gap and order, PUS TM; zero mismatches); space_packet::tests
ValidatedSPICE SPK export (types 9 and 13)spk_writer (SpkSegment,to_bytes,from_bytes,eval,et_from_label,oem_to_spk_segments)
Capability
Writer and reader of NAIF SPICE SPK kernels of types 9 (Lagrange, unequal time steps) and 13 (Hermite, unequal time steps), with the DAF layout, segment descriptors and epoch directories, and the export of a CCSDS OEM to an SPK kernel (UTC or TDB epochs to ephemeris time)
Oracle
CSPICE N0067 through spiceypy 8.3.0, leapseconds kernel naif0012.tls, pre-registered (75bc61d9) before SPICE read any Kshana kernel: SPICE's inventory of Kshana's four kernels is exact; spkpvn and spkgeo at 5 516 epochs and 690 chain evaluations equal Kshana's own evaluation bit for bit and stay inside the bar at every node (worst 0.022 of it); Kshana reads SPICE-written kernels bit-identically; for the same inputs Kshana's kernels are byte-identical to SPICE's (0 differing bytes in four files); OEM epochs convert to ephemeris time within 6.0e-8 s of str2et (bar 1e-6 s) and SPICE evaluates the OEM export inside the bar (worst 0.060). Synthetic Kepler arcs: nothing about physical accuracy. Other SPK types are not written
Tests
tests/spk_writer_spice_oracle.rs (inventory, evaluation, bit and byte identity, OEM epoch conversion; four planned mutation checks turn it red)
Writer and reader of NAIF SPICE SPK segments of type 3 (Chebyshev polynomials for position and velocity over fixed intervals, degree 0 to 27) and type 21 (extended modified difference arrays, the difference lines of JPL integrators, MAXDIM 15 to 25), from given records, in files that also hold type 9 and 13 segments; record selection and evaluation in the order of the SPICE readers. Kshana does not fit Chebyshev coefficients or produce difference lines
Oracle
CSPICE N0067 through spiceypy 8.3.0 (spkw03, spkw13 and, from the same library by its f2c symbol, spkw21; spkpvn, spkgeo), pre-registered (3f8945c7) before the code existed, the synthetic records drawn from that commit's SHA-1: for three files (type 3 degrees 0, 12 and 27 with 3 to 150 records; type 21 MAXDIM 15, 20 and 25 with 1 to 250 lines, zero to two directory entries and a centre chain; type 3, 21 and 13 in one file with comments) SPICE's inventory of Kshana's kernels is exact, Kshana's kernels are byte-identical to the ones SPICE writes from the same records, and spkpvn and spkgeo equal Kshana's reader bit for bit on Kshana's and on SPICE's files (48 876 components each way, bar 1e-12 relative). Synthetic records: nothing about physical accuracy; real JPL-produced type 21 kernels are not in the case set
Tests
tests/spk_3_21_spice_oracle.rs (inventory, evaluation both ways, byte identity); spk_writer::tests
ValidatedCCSDS Attitude Data Messages (AEM, APM)adm (AemMessage,ApmMessage,quaternion_a_to_b),ndm_xml
Capability
Writer and reader of the CCSDS 504.0-B-2 Attitude Data Messages in KVN and NDM/XML: AEM version 2.0 (and 1.0 in KVN, with its ATTITUDE_DIR, QUATERNION_TYPE and numeric rotation sequences) with quaternion, Euler-angle and spin attitude types and their derivative, angular-velocity and nutation variants, several segments; APM version 2.0 with quaternion, Euler, angular-velocity, spin, inertia and maneuver blocks; the A-to-B quaternion of a record (504.0-B-2 Annex F)
Oracle
Orekit 13.1.8 (Apache-2.0) parsers and writers, run as a separate program (xval/adm-orekit), pre-registered (2a4c44b3) before any comparison: Orekit parses every Kshana-written message (14 AEM KVN, 11 AEM XML, 4 APM KVN, 4 APM XML) and every value is inside its bar (quaternions 1e-14, angles 1e-9 deg; worst 0.054 of a bar); Kshana parses the 31 messages Orekit writes from the same inputs inside every bar; the rotation of each of 135 quaternion, Euler and spin records equals AemSegment.getAngularCoordinates within 1e-12 (worst 0.0071 of the bar). All 15 Kshana XML messages validate against the NDM/XML 4.0.0 schema (data-gated, fetched with pinned SHA-256). Finding: Orekit's APM writer refuses a maneuver without MAN_DELTA_MASS (optional in the standard), so APM maneuvers are covered in the Kshana-writes direction only. Outside the case set: AEM 1.0 */RATE types and XML, repeated APM blocks
ValidatedCCSDS Orbit Data Messages (OPM, OMM, OCM)odm (OpmMessage,OmmMessage,OcmMessage,Section,OcmBlock),ndm_xml
Capability
Writer and reader of the CCSDS 502.0-B-3 Orbit Data Messages in KVN and NDM/XML: OPM 3.0 and 2.0 (state vector, osculating Keplerian elements, spacecraft parameters, covariance, maneuvers, user-defined parameters), OMM 3.0 and 2.0 (mean elements by mean motion or semi-major axis, TLE-related parameters with BSTAR or BTERM and AGOM, covariance), and OCM 3.0 with trajectory (CARTP, CARTPV, CARTPVA; absolute or relative epochs), physical-properties, covariance (CARTP, CARTPV in LTM order), maneuver and user-defined blocks; a full-precision OMM from an SGP4 element set beside the unchanged OmmFile::to_omm_kvn
Oracle
Orekit 13.1.8 (Apache-2.0) OPM, OMM and OCM parsers and writers, run as a separate program (xval/odm-orekit), pre-registered (8efea7c1) before any comparison: 27 of the 30 Kshana-written messages (6 OPM, 5 OMM and 5 OCM cases in KVN, and XML for version 3.0) parse in Orekit with every value within 1e-14 relative and every epoch within 1e-9 s (904 values, worst 0.021 of a bar), and Kshana reads all 25 messages Orekit writes back out inside the same bars (858 values, worst 0.020); all 14 Kshana XML messages validate against the NDM/XML 4.0.0 schema (data-gated). Findings, pinned, narrowing the claim: Orekit requires ELEMENT_SET_NO (optional in 502.0-B-3), refuses COMMENT in XML user-defined parameters (the schema allows it), will not write BTERM without BSTAR (so BTERM and AGOM are covered in the Kshana-writes direction only), and adds OCM relative epochs as SI seconds, so under TDB they differ from the TDB calendar reading by about 1.9e-10 s/s (UTC, TAI and TT agree). The legacy OmmFile::to_omm_kvn output, which wraps OMM metadata in META_START/META_STOP, is refused by Orekit (recorded). Out of scope: OCM PERT and OD blocks, non-Cartesian element sets, other covariance orderings
Tests
tests/odm_orekit_oracle.rs (L1 Orekit reads 30 Kshana files, L2 Kshana reads 25 Orekit files; findings pinned and asserted to recur); odm::tests
Gravity-gradient worst-case disturbance torque ((3/2)(μ/R³)ΔI) as the peak over attitude (the validated claim), plus an RSS pointing-error budget over named 1σ contributors with the dominant term (a quadrature sum of caller-supplied numbers, outside the validated claim). Runnable from the CLI/bindings as the attitude-budget scenario kind (scenarios/attitude-budget.toml)
Oracle
Basilisk 2.9.1 GravityGradientEffector (AVSLab, ISC), an independent C++ implementation of the full-tensor gravity-gradient torque: Kshana's peak (3/2)(μ/R³)ΔI equals Basilisk's torque at the 45 deg attitude within 1e-12 relative (worst 2.1e-15) over 5 altitudes and 8 inertia tensors, and bounds Basilisk's torque over 2000 random attitudes per case; tolerance fixed before the comparison. Validated claim: the gravity-gradient torque peak. The RSS pointing budget is a quadrature sum of caller-supplied numbers with no external truth and stays outside it; the hand-coded Hipparchus full-tensor maximisation remains as supporting self-consistency. No control-loop/6-DoF/flexible-mode simulation
Tests
attitude_budget::tests (GG torque vanishes for a symmetric body, grows lower-down, linear in ΔI, RSS quadrature sum, variance-fractions-sum-to-1); tests/attitude_gg_torque_reference.rs (20 cases vs an independent…
Time-domain visibility passes (AOS/TCA/LOS, max elevation, duration) of an engine propagator over a station above an elevation mask, with interpolated rise/set crossings and total access: passes::predict_passes, the geometric scheduler (no refraction or light time). The passes scenario kind runs the apparent predictor of the row 'Apparent ground-station pass prediction with refraction and light time'
Oracle
Orekit 12.2 (CS GROUP, Apache-2.0) + Hipparchus 3.1 — an independent flight-dynamics library: ElevationDetector (Brent root-finder) + EventsLogger over an ITRF ephemeris, station as a WGS-84 TopocentricFrame. AOS/LOS/max-elevation/pass-count/total-access matched on identical orbit+station+mask+window (committed fixture; driver xval/orekit-passes)
ValidatedApparent ground-station pass prediction with refraction and light timepasses (predict_passes_apparent,apparent_look_angles,ApparentOptions,p834_tau_s_deg,p834_tau_deg,p834_theta_m_deg,p834_apparent_elevation_deg,PassesScenario)
Capability
Visibility passes (AOS, TCA, LOS, maximum apparent elevation, duration) of an SGP4 element set over a WGS-84 station on the validated Earth-orbit path (IAU 2006/2000A, optional Earth orientation parameters), with ITU-R P.834-9 tropospheric refraction (equations 9 to 14, the Recommendation's printed coefficients) and the downlink light time, each switchable; crossings solved by bisection to 1e-7 s and the culmination by golden-section search. Runnable from the CLI/bindings as the passes scenario kind (scenarios/passes.toml), both corrections on by default
Oracle
Orekit 13.1.8 (CS GROUP, Apache-2.0) with Hipparchus 4.0.3: TLEPropagator on the same element sets in a TEME Orekit builds to the stated definition (its IERS 2010 TOD rotated by GAST - GMST), ElevationDetector and ElevationExtremumDetector with and without ITURP834AtmosphericRefraction, and AngularAzEl with AngularRadioRefractionModifier for the light-time case, crossings and maximum solved by Hipparchus; 60 cases (four orbits, five sea-level stations, masks 0, 5 and 10 deg, 24 hours). Bars fixed before the run (pre-registered 896f3c2e, leg 3 of d92aef9b). Light time ignored (9777 failures) or the P.834 constant 1.728 changed to 1.828 (7883 failures) turns the test red. Scope: stations at sea level; Orekit's P.834 h*theta0^2 coefficient is 0.011380 where the Recommendation prints 0.01380 (pinned), so heights above sea level are not compared. Round 1, against Orekit's default IERS 1996 TEME, held every pass bar but missed the direction bar by the 72 mas TEME convention gap (pinned)
Tests
tests/pass_predictor_on_path_orekit_oracle.rs (60 cases, 1210 passes: AOS/LOS within 3.4e-6 s, bar 5 ms; maximum elevation within 5.5e-7 deg, bar 1e-3 deg; 8979 apparent directions within 8.8e-7 deg, bar 1e-4 deg)…
ModelledDVB-S2 adaptive coding and modulation along a LEO downlink passcontact_schedule (es_n0_db,select_modcod,fspl_db,atmospheric_loss_db,ccm_modcod,pass_capacity,AcmLinkScenario); passes::predict_passes_apparent
Capability
Downlink Es/N0 along each pass of one satellite over one station (EIRP, free-space loss, cosecant-law atmospheric loss, other losses, G/T, Boltzmann's constant, symbol rate), the most efficient DVB-S2 MODCOD that clears its ETSI EN 302 307-1 V1.4.1 Table 13 quasi-error-free threshold plus an implementation margin at each step, the pass's ACM data volume, and the constant-coding baseline that closes at the mask and the longest range. Runnable as the acm-link scenario kind. Illustrative inputs; MODELLED.
Oracle
Internal consistency: the spectral efficiency computed in closed form from Table 5a K_bch and the frame lengths agrees with the efficiencies Table 13 prints (a transcription check on the threshold table); the link-budget pieces agree with their closed forms; the selection is monotone in Es/N0 and honours the margin; ACM moves at least the constant-coding volume. The pass geometry comes from the separately VALIDATED apparent pass predictor. WHAT IS NOT: no external tool or measured link has been compared; the thresholds are the standard's ideal AWGN values, not a modem's; rain fade, pointing loss along the pass and pilots are not modelled.
ModelledMulti-satellite, multi-station ground-contact scheduling with conflict resolutioncontact_schedule (conflicts,objective,delivered,schedule_chronological,schedule_greedy,schedule_optimal,ContactScheduleScenario)
Capability
Candidate contacts from every satellite-station pass with its ACM data volume; conflicts where one antenna would serve two satellites inside its turnaround or one satellite would downlink to two stations at once; a priority-weighted delivery objective capped at each satellite's backlog; a lazy marginal-value greedy plan, a first-come baseline and an exact branch and bound on up to 26 candidates, with the greedy's optimality gap reported. Runnable as the contact-schedule scenario kind. Illustrative inputs; MODELLED.
Oracle
Internal consistency: the exact branch and bound agrees with brute-force enumeration on small random sets, every plan is checked feasible against the conflict rules, and the greedy never beats the exact optimum. WHAT IS NOT: no external scheduler or operational plan has been compared; contacts are taken whole (no splitting), the backlog is fixed at the window start, slew time beyond the fixed turnaround and onboard storage are not modelled.
ValidatedOne-way link budget (comms / link design)linkbudget
Capability
Free-space path loss, C/N₀, Eb/N₀, margin and closure over the CCSDS 401 / DSN 810-005 link equation for EIRP/G·T/range/rate/band against a required Eb/N₀. Runnable from the CLI/bindings as the link-budget scenario kind (scenarios/link-budget.toml)
Oracle
Published deep-space telecom design-control table as pinned numeric vectors: DESCANSO / J. H. Yuen (ed.), Deep Space Telecommunications Systems Engineering, JPL Pub 82-76, Table 1-1 (Galileo X-band) — kshana reassembles the table line-items and reproduces its published end-to-end L_fs 290.54 dB and Pr/N0 54.6 dB-Hz; free-space loss also checked vs ITU-R P.525
ModelledFrugal cost-per-coverage / ROI framingfrugal (over walker)
Capability
Cost-per-percent-coverage + coverage-per-euro ROI over the constellation sizing engine; per-satellite cost is a caller-sourced low/nominal/high bracket (no fabricated prices)
Oracle
Closed-form cost arithmetic vs hand-derived values; an economic FRAMING of a modelled coverage figure, not a quote or validated cost model
Tests
frugal::tests (hand-derived cost-per-coverage 48/96=0.5, ROI ratio 2.667, bracket-ordering + zero-coverage guards)
ModelledDetection-miss integrity impact (context-aware HPL/VPL vs alert limit)integrity_impact (over raim)
Capability
Maps an undetected spoof/jam bias to effective error → Stanford region (available/unavailable/MI/HMI) against context-specific HAL/VAL (open-sky vs urban)
Oracle
Composes the externally-validated RAIM Stanford classification (raim::classify_stanford); the detection-miss→AL mapping itself is modelled, not a certified integrity allocation
Tests
integrity_impact::tests (same miss flips Available→MI→HMI as the context tightens; conservative-PL→Unavailable; per-axis HMI; input guards)
Bracketed (best/nominal/conservative) cold-atom-interferometer performance — bias instability, velocity/angle random walk, scale-factor stability, interrogation-limited sample rate, fringe-ambiguity dynamic range — each citation-traceable; feeds QuantumNavBudget without modelling hardware
Oracle
Internal consistency: the cited VRW bracket cross-checked against CaiAccelerometer::accel_asd physics + raw dynamic range computed from the fringe-ambiguity limit; numbers are MODELLED literature-survey brackets (needs_source_confirmation), no device validated, no validation halo
Tests
inertial::cai_params::tests (physics VRW lands inside the cited VRW bracket at all 3 levels; raw fringe-ambiguity range computed from k_eff·T²; conservative budget drifts more than best; every bracket sourced +…
Independent numpy re-implementation of the four-region rule (scripts/gen_tib_scorer_reference.py) on the identical sample set; method Cited from the Stanford–ESA integrity diagram (Tossaint et al., ION GNSS 2007) and RTCA DO-229 WAAS MOPS. NOT an ExternalDataset accuracy oracle — the benchmark is honesty-immune.
Tests
src/benchmark/stanford.rs::tests (four regions + inclusive boundaries + |error| + PL>AL unavailability); tests/tib_scorer_reference.rs (classification counts vs independent numpy on a fixed synthetic set)
Representative parametric fault generators (Modelled); the undetectable set (symmetric delay, replay-within-freshness) is Cited from Mizrahi RFC 7384 and Narula & Humphreys (IEEE JSTSP 2018). No external oracle — a fault catalog makes no measured claim.
TIB undetectable-absorption verdict (symmetric/replay must be absorbed by the PL, never reported detected)
Oracle
Property check that the Verdict enum structurally cannot report an undetectable fault as detected (Mizrahi RFC 7384); absorbed ⟺ PL ≥ offset. No accuracy oracle.
Tests
src/benchmark/scorecard.rs::tests (reference PL absorbs; broken PL unabsorbed-never-detected; detectable coverage verdict; the honesty property that no coverage/detection verdict is reachable for an undetectable…
Partner-ownedSpacecraft bus engineering (AOCS/thermal/structures/propulsion/power)
Capability
Not provided — Kshana is a navigation-performance simulator, not a bus house
Ashby & Patla 2024, 'A Relativistic Framework to Estimate Clock Rates on the Moon', Astronomical Journal 167:149 (NIST; basis for the IAU/IAG LunaNet LTC): Moon-surface self-potential L_m = 3.13881e-11 and secular total 56.02 µs/day matched as published numeric values; geocentric Moon speed cross-checked vs JPL DE440 (de440s.bsp via SPICE)
Tests
lunar_time::tests (closed-form rate; round-trip); tests/lunar_coordinate_time_reference.rs (LTC self-potential & secular-rate terms vs published Ashby & Patla 2024 values + geocentric Moon speed vs JPL DE440)
The hybrid-optical-rf report states the link configuration it actually ran at (carrier wavelength, transmit and receive aperture, range, pulse width, integration time, efficiencies and losses, defaults resolved) and carries a units block giving the unit and provenance class of every quantity a paper is likely to quote, including the handoff covariance traces in square metres
Oracle
No external oracle, and none is possible: this is self-description, not a measurement. It is recorded because the absence of it was a defect -- P5 had to quote the carrier wavelength and transmit aperture from source defaults, and to infer that a bare variance was in square metres from an internal consistency check. The inference was correct, which is precisely why it mattered: nothing would have caught it being wrong
Tests
hybrid_integrity::tests (the resolved configuration is echoed for defaults and for overrides, without round-tripping the wavelength through metres; every units entry carries both a unit and a provenance class; every…
ModelledHybrid optical/RF link availability on the RF sidehybrid_integrity,linkbudget,jamming
Capability
The hybrid-optical-rf report states an RF link availability, composed from quantities the engine already computes rather than from an imported climatology, and states the composition as a named rule: A_rf = I_closure * I_track, where I_closure is the closure verdict of the one-way CCSDS-401 / DSN-810-005 link budget (linkbudget::link_budget, Eb/N0 margin >= 0) at the scenario's own range and I_track is the tracking-loop verdict (jamming::lock_status) on the C/N0 that SAME budget returned. Each factor carries its input's provenance class through to the output, and the factors deliberately NOT in the product -- geometric visibility, interference denial, and an RF outage climatology -- are named with the reason rather than silently set to 1. The resolved RF leg is echoed as its own rf_link_configuration block, and the continuous figures beside the indicator (link margin, C/N0 margin, the closed-form closure and tracking ranges, and the range utilisation) are the range inversions of the same budget. The report states in full, on the block and in the units entry a reader lands on, that this figure is MARGIN/GEOMETRY-LIMITED and DETERMINISTIC while optical availability is WEATHER/CLIMATOLOGY-LIMITED and probabilistic, so the two can never be quoted as a comparable pair of percentages
Oracle
No external oracle, and ExternalDataset is declined. The two inputs are an Eb/N0 margin and a C/N0 threshold crossing, both functions of a MODELLED EIRP, figure of merit and lumped loss allocation; there is no measured availability record for an Earth-Moon optical/RF hybrid service to check the composed figure against. The checks that exist are internal and are stated as such: the composition is recomputed from the report's own published margins rather than from the emitter's internals, and the closure range is verified by re-running the link equation AT it and requiring the margin to vanish -- an inversion round trip, not an independent implementation. The honest limit of the figure is recorded on the report itself: it is a deterministic 0/1 indicator, not a probability, because nothing in this engine measures an RF link-outage distribution at this band and geometry. That missing input is named in rf_availability.factors_not_included rather than invented, and supplying it is what a probabilistic RF availability would need
Tests
hybrid_integrity::tests (the availability equals the product of the two indicators recomputed from the report's own margins, and those margins are themselves reassembled from the report's own EIRP, free-space loss…
The hybrid-optical-rf report emits the optical-versus-RF ranging ratio with the ONE configuration both legs were evaluated at carried in the same object: the same one-way path, the same range, and the same accumulation time. The optical leg is the engine's photon-limited ToA CRLB on the one-way photon count; the RF leg is the engine's DLL early-late thermal code-tracking jitter at the C/N0 the engine's own link budget returns for that same one-way range. The loop noise bandwidth is derived, not chosen: B_L = 1/(2*integration_s) puts the RF leg at exactly the optical accumulation time, and if a caller overrides it so the two averaging times disagree the ratio is REFUSED -- null, with the mismatch and both times named -- rather than quoted at two operating points. No ratio is formed against the scenario's CHOSEN parametric rf_position_sigma_m, and the refusal says why: that input carries no configuration at all. The released two-way optical headline and the exact factor bridging it to the one-way comparison leg are emitted beside the ratio, so the report cannot be read as carrying two disagreeing optical sigmas
Oracle
No external oracle for the RATIO, and ExternalDataset is declined deliberately. The underlying RF chain does have one -- tests/validate_p5_rf_ranging_precision.rs checks linkbudget::link_budget and navsignal::dll_code_jitter_chips against an independent Python/NumPy fixture and against the hardcoded Kaplan & Hegarty worked value (2.814e-3 chip / 0.825 m at 45 dB-Hz) -- and the optical CRLB has its own closed-form row. Claiming either anchor for this row would be borrowed validation, which the matrix invariants exist to prevent: the quantity here is a RATIO at a configuration with a MODELLED EIRP, figure of merit, transmit power and aperture, and no measured optical-versus-RF ranging comparison at a common operating point exists to check it against. What is checked internally is the thing the ratio can actually get wrong: that the two legs sit at one operating point. That is measured by scaling the common accumulation time and requiring both legs to move by the same square-root law and the ratio to stay put, and it is enforced by refusing the ratio outright when the averaging times disagree. Both legs are thermal/shot-noise bounds and both exclude media delay, clock error and ambiguity, so the exclusion is identical on each side; the report says so rather than leaving it inferred
Tests
hybrid_integrity::tests (the ratio equals the two emitted legs divided, to 1e-15 relative, and its reciprocal and decibel forms agree; each leg's range and time sigma are related by exactly c; the emitted common…
The link-budget report states every absolute constant its own margin was computed from -- the carrier frequency the free-space loss used, the EIRP, the figure of merit, the lumped loss and the Boltzmann term -- plus the required G/T at which the margin is zero, the link constant (EIRP - losses - required Eb/N0) that is the only combination a published rate/gain table can ever fix, and, when a caller states a system noise temperature, the receive antenna gain that figure of merit implies
Oracle
No external oracle: this is self-description over an equation already validated against a published design-control table (see the one-way link budget row). It is recorded because the absence of it was a measured defect -- reproducing the released d1_rate_gain_beamwidth.csv required back-solving one effective constant to 0.0034 dB from all thirty rows, and the engine now both names that combination and proves, by test, that no released table could ever have separated its three components
Tests
linkbudget::tests (across three bands, four decades of range and both closure verdicts, the margin and the free-space loss under it are recomputed from the report alone to better than 1e-9 dB; the required G/T…
ModelledPer-clock-class lunar time crossover tablelunar_time_budget
Capability
One lunar-time-budget run emits a clock-vs-frame crossover row per clock class against a single shared frame term, so the clock is the only variable in the comparison; each row carries the crossover reached two ways -- bisected from the general power-law time-error curve and inverted algebraically from the row's dominant noise type -- with the relative difference between them
Oracle
Two different in-codebase computations of the same quantity, not one restated: bisection on the IEEE-1139 power-law curve knows nothing about noise type, while the closed form inverts the dominant type algebraically, so a misclassified noise type or a bad bracket shows as a non-tiny relative difference. The four values also reproduce the paper's published table to half its last printed digit -- but that table was itself reconstructed from the same closed form, so it is a reproducibility check and not an external oracle, and the frame term it is measured against is a Modelled allocation
Tests
lunar_time_budget::tests and lunar_time_budget_scenario::tests (all four classes present in one run and in a requested subset order; every row agrees with its own closed form to better than 1e-12 relative; scaling the…
ValidatedJoint UT1 and polar-motion error over a common row setframe_eop,realtime_frame_eop
Capability
One table reports the UT1 prediction error, the polar-motion pole error and their quadrature combination at the Moon over an IDENTICAL epoch set per horizon, emitting the epochs each component was measured at. Separately, the scenario names whichever EOP input is in force and decomposes its row census, and always emits the predicted-versus-final horizon table with an explicit statement of why it is empty when it is
Oracle
astropy 8.0.1 utils.iers (BSD-3) reading the IERS finals2000A product with its own Bulletin-B-else-A rule, and pyerfa 2.0.1.5 (SOFA) full celestial-to-terrestrial rotations (c2i06a, sp00, pom00, c2tcio; the Earth rotation angle per IERS 2010 eq. 5.15 in exact arithmetic), run as separate programs on verbatim rows of the frozen 2026-09-30 finals2000A.all; pre-registered (3d43ba52) with amendments A1 and A2 fixing only the oracle numerical precision (both disclosed with the runs they followed): every Table 3 epoch list equal, UT1 and pole statistics within 6e-16, the combination within 1.9e-7, 12/12 predicted-versus-final cases. The combination check cannot resolve pole-term errors below about 1e-4 relative because UT1 dominates it; pole accuracy rests on the 1e-9 pole-statistic check. The epoch-set and fallback rules are the generator mirroring Kshana definitions, and the empty-table statuses are design behaviour with no external truth. Magnitudes: IERS Annual Report 2019 Table 3a realised prediction errors, Kshana/IERS 1.001 to 1.100 at 1 to 90 days (bar [0.8, 1.25], Measured). Row census: astropy (offline-input row). Round 1 (kept as a regression): against the accuracy formula printed in 178 Bulletin A issues the pole agrees within 1.5x and UT1 does not, which tests the IERS formula against its realised error
Tests
frame_eop::tests and realtime_frame_eop::tests (all three components carry equal, elementwise-identical, strictly ascending epoch sets over both real IERS extracts; the joint set collapses to the intersection when the…
ValidatedOffline default Earth-orientation input is a real IERS productrealtime_frame_eop,eop
Capability
The realtime-frame-eop runtime default is the library own embedded copy of a verbatim IERS finals2000A extract (tools/finals2000A_20260930.txt: 174 consecutive daily rows, MJD 61224-61397, cut from the finals2000A.all of 2026-09-30) carrying all three row vintages the format defines, classified by the IERS I/P flags (eop::row_vintage): 30 Bulletin B finals, 54 rapid measured rows (flag I without a Bulletin B block) and 90 Bulletin A predictions (flag P). A bare run with no file argument and no network emits a populated per-horizon table, the census rows = final_rows + rapid_rows + prediction_rows (174 = 30 + 54 + 90) and predicted_rows.n = 90 over MJD 61308-61397; the as-issued cutoff of the predicted-versus-final comparisons is the last measured row. The earlier 32-row extract (tools/finals2000A_2026.txt, every row flag I) and the five-row final-only excerpt stay shipped and exercised: on them predicted_rows.n is 0, a property of the input file and not a parser outcome
Oracle
astropy 8.0.1 astropy.utils.iers.IERS_A (BSD-3-Clause), reading the same bytes with its own reader, and the IERS Bulletin A Vol. XXXIX No. 039 table, copied number for number; pre-registered (74b3c1a0) with six exact criteria before astropy was run. All six hold: 30 final, 54 rapid and 90 predicted rows with equal MJD lists, every predicted row flag P, the default-run census equal, and the predictions equal to Bulletin A within the last printed digit. Disclosed: the cut was chosen with a plain column slice of the flags and the Bulletin B block, which showed 30/54/90, before the commit; the extract is a verbatim slice this repository cut from an IERS product (SHA-256 in the fixture NOTICE). The validated claim is the vintage classification and census of the shipped default input and its predicted rows; arbitrary user inputs are classified by the same flag rule but are not part of the comparison. 0.30 round 1 (superseded): the earlier default called 12 flag-I rows Bulletin A predictions; the engine now classifies by flag. REVISIONS: the default-run cells that moved are listed old-to-new in docs/revisions/M035-default-eop-recut-cell-changes.md (predicted_rows 12 -> 90, the pole floor 0.06776 -> 0.05550 mas, the Earth-orientation term 14.01601 -> 14.01583 m, the total 20.09759 -> 20.09746 m and 67.03834 -> 67.03790 ns, the Table 2 rows of tests/golden/realtime-frame-eop.csv, and the operational-predictor agreement figures); earlier ones in docs/revisions/G12-default-eop-cell-changes.md
Tests
tests/embedded_eop_vintage_astropy_preregistered.rs::vintage_of_every_row_matches_astropy_and_predictions_match_bulletin_a (six exact criteria: the extract bytes; 174 rows; per-vintage counts and full MJD lists…
ModelledCapture footprint against altitude and beamwidthantenna
Capability
Two-axis sweep of the pattern-weighted, altitude-limited surface capture footprint over transmitter altitude and dish diameter, emitted one row per operating point with the half-power beamwidth each diameter implies, and limb capture reported as a THRESHOLD -- per row the limb J/S, its margin and the transmit power that would close it; per grid the located crossings, or an explicit statement that the limb is not reached anywhere on the grid, with the shortfall
Oracle
Set inclusion: the jammer-to-signal ratio enters as a uniform decibel offset, so the captured set at a higher transmit power contains the set at a lower one and the fraction can only rise -- a property the Airy pattern does NOT give in beamwidth or altitude, where the captured region breaks into rings and the fraction is genuinely non-monotone. No published table gives the captured disk fraction of a lunar orbital transmitter against altitude and beamwidth, so there is nothing external to check these cells against; the pattern underneath is separately Validated against published Bessel values and keeps its own row
Tests
antenna::tests and attack_surface::tests (the grid is complete and every captured fraction lies in [0,1]; captured fraction is non-decreasing in transmit power at every node; it is NOT monotone in beamwidth or…
The lunar-time-budget scenario publishes its array-valued outputs as a long-form (grid index, averaging time, term) table alongside the report, so the seven per-term x(tau) curves and the root-sum-square total are engine output rather than something a reader rebuilds from the method section
Oracle
Self-consistency only: the published total is checked against the root-sum-square of the terms in the same file, and the emitted curve reproduces the released p3_time_budget_curve.csv over all 57 points. Both checks share this engine's own term definitions, so neither is independent. The clock term rests on published clock specifications, but the link, frame, relativistic and ephemeris floor MAGNITUDES are documented budget allocations with no external oracle
Tests
lunar_time_budget_scenario::tests (all 57 averaging times x 8 terms present, the grid index runs 0..=56 and carries 8 rows each; the total row equals the root-sum-square of the seven terms beside it at every tau; the…
ModelledLunar station time laboratory: miniRAFS holdover against LTClunar_time_lab (over clock_specs,lunar_time_budget,lunar_time)
Capability
The optional [time_lab] block of the lunar-time-budget scenario models a lunar reference station clock (default the miniRAFS class, or any named class, or a device card fitted to stated stability points) steered to Coordinated Lunar Time and then free-running or re-steered every N seconds: the time-error curve x(t) as the root-sum-square of the clock power-law term, a constant steer residual (ILLUSTRATIVE), a constant LTC realisation residual and an opt-in station-height rate term; the time after a steer to 1, 10 and 100 ns (or stated thresholds) by bisection beside the closed-form inverse for the dominant noise type, and the dominant term at each stated horizon, all with units
Oracle
Internal consistency only: the bisected time to threshold is checked against the closed-form inverse of the clock's dominant noise type, and the free-running clock term against the budget's own clock term. The clock classes rest on published one-day specifications (their own rows); the LTC rate is not re-derived here and keeps its own VALIDATED row (Lunar coordinate time); the steer residual is ILLUSTRATIVE, the realisation residual is a budget allocation and no thermal model is included. The nanosecond-level target described in the ICG WG-B Lunar PNT material of June 2024 is quoted as a source statement, not adopted as a claim; nothing here is a statement about any real station clock
Tests
lunar_time_lab::tests (free_run_clock_term_is_the_budget_clock_term: with steering off the clock term equals the budget's clock term to the bit for all four classes; bisection_matches_the_closed_form within 1e-9…
ModelledLunar geodetic VLBIlunar_vlbi
Capability
Near-field VLBI delay for an Earth baseline observing a lunar beacon + partials
Oracle
Plane-wave delta_dor (same-codebase) in the far-field limit; finite-difference partials. 0.30 external comparison, a finding (stays MODELLED): against ANISE 0.10 converged light times through DE440 (75 delays on 2024-01-01 over three DSN baselines, generated by xval/anise-lunar-od without Kshana code), no delay is within the 1 ps tolerance. The largest gap is 24 µs, from the analytic Moon centre (about 200 km from DE440), the single instantaneous geometry with no light-time iteration (3.6 µs) and the dropped polar motion; beacon partials differ by up to 2.9e-3 relative against 1e-6. The module-doc claim that the frame approximations are below the model fidelity does not hold at this level. 0.30 round 2: the kernel path (lunar_vlbi::KernelGeometry) is validated against NAIF SPICE light times in its own row, Lunar geodetic VLBI, kernel path; the ANISE comparison of that path (75/75 within 1 ps, worst 0.12 ps, tests/lunar_vlbi_anise_oracle.rs::kernel_delay_matches_the_anise_light_time_difference) is a kernel-evaluation cross-check only, because the light-time solve in that harness is this project own. This row (the analytic default path) stays MODELLED with the finding above
Near-field VLBI delay for an Earth baseline observing a lunar beacon, with beacon and station partials, on the kernel path (lunar_vlbi::KernelGeometry): DE440 Earth and Moon, DE440 lunar principal axes, ITRF93 Earth orientation with UT1 and polar motion, each light time converged in the barycentric frame, partials with the light-time factor 1/(c - u.V). Runnable as the lunar-vlbi scenario with planetary_kernel_path, earth_orientation_kernel_path and moon_orientation_kernel_path set; the validated outputs are the kernel-path geometric_delay_s series and epoch_partials. On the kernel path delay_s adds the analytic differenced Shapiro term, which is not compared, and delay_rate and near_field_correction_us are not compared
Oracle
NAIF SPICE Toolkit N0067 (spiceypy 8.2.0, MIT) converged Newtonian light times (spkcpt/spkcpo, CN) from a beacon fixed in MOON_PA_DE440 to the DSN stations DSS-14, DSS-43 and DSS-63 from NAIF station kernel earthstns_itrf93_260814.bsp, with de440s, moon_pa_de440_200625, moon_de440_250416.tf, earth_latest_high_prec and naif0012; partials are five-point central differences of SPICE light times, not Kshana 1/(c - u.V) formula. Pre-registered (22caf18f, 2024 epochs) with bars of 1 ps on every delay and 1e-6 relative on every partial. That first run failed (17/75 delays within 1 ps, worst 1.03e-11 s): SPICE carries ET as one double, whose spacing at 2024 rounds the emission epoch by up to 6e-8 s, about 6 ps of light time, and the residual matches that rounding at 0.99999 correlation (pinned; the reviewer reproduced it independently). Amendment 8196cc55, written after that result and before its fixture existed, moved the same oracle and bars to 25 hourly epochs near J2000, where the ET spacing is at most 1.5e-11 s: 75/75 delays within 1 ps (worst 6.2e-14 s), 75/75 beacon partials (worst 7.4e-9), 150/150 station partials (worst 3.8e-11). Outside the claim: the analytic default path (series Moon and mean lunar frame, 24 µs from ANISE through DE440, the Lunar geodetic VLBI row), the differenced Shapiro term the scenario adds analytically, media, and the barycentric-to-geocentric scale
Tests
tests/lunar_vlbi_spice_oracle.rs::kernel_delay_and_partials_match_spice_light_times_near_j2000 (75 delays at 25 epochs near J2000 over DSS-14, DSS-43 and DSS-63: delays within 1 ps, worst 0.06 ps; body-frame beacon…
ValidatedNAIF kernel reader: DAF container, SPK type 2 and binary PCK type 2naif_kernel (DafFile,SpkKernel,PckKernel); ephem_provider (KernelEphemeris)
Capability
The engine's own pure-Rust reader of NAIF binary kernels (naif_kernel): the DAF container (both byte orders, linked summary records), SPK type-2 Chebyshev position segments chained through their centres to a lowest common ancestor, and binary PCK type-2 Euler-angle segments turned into a J2000-to-body rotation, with epochs taken as two-part ET so a sub-microsecond offset keeps its precision. It is what KernelEphemeris (ephem_provider) and the kernel paths of lunar_vlbi, lunar_llr and lunar_vlbi_fim read DE440 through. The claim is that it reads JPL kernels correctly; it is not a claim that DE440, or the engine's analytic Sun and Moon series, is accurate
Oracle
Two independent kernel readers on the same NAIF files: NAIF SPICE Toolkit N0067 (CSPICE through spiceypy 8.2.0, MIT; spkezr and pxform) and ANISE 0.10.6 (Nyx Space, MPL-2.0; Almanac.translate and Almanac.rotate), both run as separate programs on de440s.bsp and moon_pa_de440_200625.bpc. Pre-registered (1e283534) before any fixture or oracle value existed: the grid seed is the first 16 hex digits of that commit's hash, so neither the 200 epochs nor the body pairs could be chosen with the result in view; each epoch also carries the Moon and the Sun relative to the Earth. Bars fixed in advance from double-precision Chebyshev evaluation (at most 15 coefficients, n^2 u = 2.5e-14 per evaluation, two evaluations and up to two hops per side): position 1e-13 R + 1e-5 m, velocity 1e-13 V + 1e-10 m/s, R and V the larger magnitude relative to the lowest common ancestor; rotation element 1e-13 W + 1e-15, W the coefficient sum of the covering record; ANISE adds 2 ns of relative motion for its nanosecond epochs. Result: every one of 600 states and 200 rotations inside its bar against both oracles, on the cut and on the full kernels; worst 6.0e-3 of the bar (position), 5.9e-3 (velocity), 3.7e-3 (rotation). Mutation: evaluating each record at -s fails 3597 comparisons. Disclosed: an earlier 25-epoch check against SPICE (4.6e-5 m, tests/naif_kernel_reader_check.rs) was seen before this row and is not its evidence; the fixture generator's first run aborted on an empty grid interval before producing any value and was corrected. Both sides read DE440, so agreement validates the reader, not the ephemeris
Tests
tests/naif_reader_spice_oracle.rs::reader_matches_spice_and_anise_on_the_post_registration_grid (600 de440s.bsp states and 200 MOON_PA_DE440 rotations at 200 epochs 1849-2150 and random body pairs drawn from the…
ModelledEarth-GNSS at lunar distanceearth_gnss_lunar
Capability
Earth-GNSS reception at lunar distance: the weak-signal layer P7 Table 1 names and the engine had no model for. Per satellite and per epoch it computes the off-boresight angle against the transmitter nadir, whether the Earth occults the straight path, the slant range and its free-space loss, the transmit gain at that angle, the received power and the carrier-to-noise density; it then aggregates only the links clearing a tracking threshold, and sweeps a full constellation revolution so the answer is a DISTRIBUTION rather than one snapshot. Three facts drive the result and the report states each: the Earth subtends 13.90 deg from the constellation radius so the beam PEAK is geometrically unavailable to any lunar-bound ray; what remains is the main-lobe edge and the sidelobes, the regime LuGRE operated in at the Moon in 2025; and the L1 path loss is about 208 dB, some 25 dB more than a terrestrial user pays. MEASURED on the bundled 24-satellite GPS-class geometry over 64 epochs of one orbital period: SIGNAL availability 0.375 with a best link of 30.93 dB-Hz, and FIX availability 0.000 - at most two simultaneous trackable links, never the four a position needs. The separation of those two availabilities is the point, and the report refuses to let them be confused: Earth-GNSS at lunar distance is a TIMING-grade layer, not a position-grade one, and a layered-resilience prior must take the fix availability. The conditioning is emitted for the same reason - every visible satellite lies inside a cone a couple of degrees wide as seen from the Moon, so the lines of sight are near-parallel by construction. Runnable as the earth-gnss-lunar kind Restated per D7: the claim at lunar distance is the RELATIVE carrier-to-noise density between satellites at one epoch, from measured two-dimensional transmit patterns (NAVCEN Block IIR/IIR-M), plus geometric visibility; absolute C/N0 is not claimed (Parker et al. report a common 7 to 12 dB loss).
Oracle
Closed-form identities and internal consistency; NO external oracle is claimed. The limb half angle is checked against asin(R_earth/r_orbit) and the geometry against it; the free-space loss, the DOP kernel and the SGP4 propagation this composes are each externally validated in their own rows and this row does not borrow their status. WHY MODELLED, and it is the transmit pattern: the gain at angle is a uniformly illuminated circular aperture, the Airy pattern, while a real GPS L1 antenna is a twelve-element helical array with a shaped main lobe peaked off-boresight to even out power across the Earth disc and sidelobes that are not Airy. Published measured patterns exist and are not vendored here, so orderings and orders of magnitude carry, while the dB of any single satellite does not. A FIRST IMPLEMENTATION OF THIS ROW WAS WRONG IN A WAY WORTH RECORDING: the Airy expression is valid only in the forward hemisphere, its argument goes as sin(theta), and at theta near 180 deg it wraps around and returns FULL BORESIGHT GAIN behind the aperture - so the two strongest links in the first report were satellites pointing their antennas away from the Moon, and they were the only two that cleared the threshold. Behind the aperture the model now applies a flat back-lobe floor, stated as a bound rather than a fabricated pattern shape. Deliberately absent, each making the budget OPTIMISTIC: no ionospheric or tropospheric loss on the limb-grazing rays, no polarisation, pointing or implementation loss, and a spherical Earth with no refractive extension. The Moon position is an INPUT, not an ephemeris lookup, because the quantity under test is the link and the beam geometry. Upgrading this row to Validated needs a measured transmit pattern and LuGRE normal points to check against; neither is in the repository and neither is invented here. 0.30 round 2, a finding (stays MODELLED): a measured-input link path (real SP3 positions, per-block transmit power, tabulated transmit and receive patterns, acquisition and tracking hysteresis) was checked against Montenbruck et al. 2023 Table 3 for GENESIS at 6000 km (pre-registered 758df4eb): zenith-antenna means agree on 5 of 6 within one dB of the paper sensitivity; nadir-antenna means are 1.0 to 2.4 satellites higher than printed, plausibly because the paper used azimuth-dependent GPS side-lobe patterns that are not published. The lunar-distance figures of this row still use the Airy stand-in. 0.30+ D7, a finding (stays MODELLED): the restated quantity, the C/N0 DIFFERENCE between Block IIR/IIR-M satellites tracked at one epoch, predicted with the NAVCEN per-SVN measured L1 patterns (two-dimensional, yaw-steering frame) against the LuGRE flight C/N0 (pre-registered 7425dc34: 99 % visible, pair RMS <= 3.0 dB, median <= 2.0 dB, at least 50 records and 30 pairs) has too little data: the 26 qualifying transit and lunar-orbit epochs hold 15 tracked IIR/IIR-M records (mostly IIF and III are tracked) and one pair. All 15 are predicted visible and the single pair's residual is -0.06 dB; nothing is claimed
Tests
earth_gnss_lunar::tests (11 lib tests: the Earth-limb half angle against its closed form asin(R/r) with a further assertion that EVERY un-occulted link lies outside that cone, so the occultation test and the limb angle…
ModelledLunar orbit determination: Earth-GNSS against ground radiometriclunar_infra::{orbit,gnss,od}
Capability
Orbit determination in a near-rectilinear halo orbit (or an elliptical frozen orbit) from two independent data types: Earth-GNSS pseudoranges received on board (the link rule of the Earth-GNSS-at-lunar-distance row, plus Moon occultation) and ground two-way range and range-rate from three Deep Space Network stations. Each is a one-step batch least squares about the reference orbit with its formal covariance; a seeded Monte Carlo of the difference between the two estimates tests whether the two covariances explain it (q = d' (P_G + P_R)^-1 d, chi-square with six degrees of freedom when the models match) and what an unmodelled acceleration does to the statistic.
Oracle
An independent scipy/numpy program integrates the same variational equations (bar O1). Nothing is compared with a measurement of a lunar orbit. The oracle checks the numerics of stated formulas, not the physics, so the row stays MODELLED. Restricted three-body model: no Sun, no lunar harmonics, the Earth rotation axis along the orbit normal, ILLUSTRATIVE receiver and station inputs.
Tests
tests/lunar_infra_od_consistency.rs (bar O1: the state-transition matrix chained over three days and over one period equals scipy DOP853 to a relative 1e-6; bar C1: mean q and the in-95 % fraction; bar C2: sample over…
ModelledLunar carrier-phase ambiguity resolution: surface and orbiterlunar_infra::phase; lambda
Capability
Between-receiver single-differenced Earth-GNSS carrier phase and code at a lunar surface station and an orbiter, with the per-epoch clock difference profiled out, one ambiguity per continuous pass and the longest pass as the gauge. The float ambiguity covariance is a Schur complement; LAMBDA resolves the integers, the bootstrapped rate is reported beside a seeded Monte Carlo integer-least-squares success rate and the float and fixed position sigmas. Reported as found: at the registered 12 dBi the double-difference problem is not formable (amendment 1 moved the default to 25 dBi after availability counts only were seen), and at 25 dBi integer fixing is NOT achieved - out of reach (a few percent at most) although the float solution improves with the arc.
Oracle
An independent numpy solve of the same normal equations from the exported rows (Q3) and a brute-force integer enumeration (Q4). No GNSS or lunar carrier-phase measurement is compared with. MODELLED.
Tests
tests/lunar_infra_carrier_phase.rs (bar Q1: Monte Carlo success at least the bootstrap bound minus three standard errors; Q2: no fall with arc length beyond three standard errors; Q4: LAMBDA equals brute-force…
ModelledLunar frame co-estimation: surface stations and orbiterslunar_infra::frame; lunar_datum; lunar_identifiability; lunar_llr_geometry
Capability
Fisher information of the seven-parameter lunar frame datum from Earth-station laser ranging to the five reflectors alone and with orbiter-to-surface-station ranges added: the information matrix, rank, Cramer-Rao sigmas, the degeneracy metric of the lunar identifiability module and seeded random station sets. The orbit is held known in the lunar axes (its error is carried by the orbiter range sigma).
Oracle
An independent numpy finite difference of the forward model (F2). Nothing is compared with real lunar frame solutions. The factor is conditional on the assumed known orbit and the 1 m range sigma. MODELLED.
Tests
tests/lunar_infra_studies_3_5.rs (bar F1: the CRLB trace does not rise and the smallest eigenvalue does not fall with orbiter links, in the example and in 20 seeded random sets; bar F2: the range partials equal an…
ModelledSpace-tie frame-tie precision: co-located GNSS, VLBI and laser ranginglunar_infra::tie; fim::crlb_srif
Capability
Cramer-Rao precision of the inter-technique translation, scale and rotation offsets between GNSS, VLBI and laser-ranging frames when one spacecraft carries all three, with the spacecraft position eliminated at every epoch and nine local-tie nuisance parameters carrying a prior, swept over the local-tie accuracy. Covariances by the square-root route (normal equations lose the digits when the techniques are near perfect).
Oracle
An independent numpy stacked system with the position unknowns eliminated by explicit Schur complements (T3). No GRASP, Genesis or other mission data are compared with. MODELLED.
Tests
tests/lunar_infra_studies_3_5.rs (bar T1: translation sigmas do not fall as the local tie loosens; T2: with technique sigmas scaled by 1e-6 each pair's translation sigma is within 1 % of sqrt(2) times the local-tie…
ModelledLunar Love-number and libration sensitivitylunar_infra::geodesy; lunar_od::degree2_tide_accel; lunar_time
Capability
Fisher information of h2, k2, l2 and the physical libration amplitudes from laser ranging (radial and horizontal tidal displacement, reflector rotation), orbiter tracking (the degree-2 k2 acceleration integrated along the line of sight, trend removed), a surface gravimeter (gravimetric factor 1 + h2 - 1.5 k2) and a surface clock (fractional rate (1 + k2 - h2) W2/c2, tied to the lunar-time constants). h2 and l2 nominal values are linearisation points; the study says what is sensitive to what, not what the numbers are.
Oracle
An independent numpy finite-difference assembly (G2) and closed-form evaluation (G4). The k2 sensitivity of the orbiter family is an upper-end figure (the orbit state absorbs only a constant and a trend here). No lunar tide or libration measurement is compared with; k2 is quoted from Lemoine et al. 2014 as a nominal. MODELLED.
Tests
tests/lunar_infra_studies_3_5.rs (bar G1: analytic partials equal a central finite difference of each forward model to 1e-8; G2: the 5x5 information equals an independent numpy assembly to 1e-6; G4: the clock tide…
Beacon-augmented dilution of precision for a lunar surface user: the visible-satellite line-of-sight rows and the visible-surface-beacon ranging rows are concatenated into one design matrix and evaluated through the shared DOP kernel, and the resulting DOP is mapped to a realised 1-sigma accuracy IN METRES through a per-beacon user-equivalent ranging error assembled as the root-sum-square of clock-synchronisation, multipath and survey terms. Beacon visibility is the airless-Moon two-height geometric horizon, which has no refractive extension and is therefore exact rather than approximate. Runnable as the lunar-beacon scenario kind, which reports satellites alone, satellites plus beacons, and a larger constellation as the competing route to the same geometry. MEASURED on the bundled golden geometry (user at -80 deg, three surveyed beacons, six-satellite illustrative LCNS at t=0, 5 deg mask): 5 visible satellites give PDOP 9.6941 and a 3-D 1-sigma of 11.222 m; adding the beacons that actually clear the horizon gives PDOP 4.1160 and 4.765 m, a factor of 2.355; the 24-satellite service instead gives PDOP 2.4226 and 2.804 m, a factor of 4.002. The report prints the VISIBLE beacon count rather than the configured one, and on this geometry only ONE of the three beacons clears the horizon: the two flanking sites lie about 333 km from the user against an 86 km horizon for a 2 m antenna, so a reader is never left to assume all three contributed
Oracle
The DOP arithmetic is the crate::orbit::dop kernel, separately externally validated against gnss_lib_py in tests/dop_reference.rs; this row does not re-borrow that row's status. The beacon-visibility horizon is the L01 closed form, and sigma = DOP x sigma_URE is the standard GNSS relation (Kaplan and Hegarty, Understanding GPS/GNSS, section 7). Within this row the augmentation itself is checked against an independent in-repo DOP path and pinned to a committed golden. WHY MODELLED: the constellation design, the beacon placement, the antenna heights and all three error-budget magnitudes are illustrative inputs rather than a fielded survey or a measured link, so the reported metres are a property of a chosen scenario and not of any deployed service. NOTE ON PROVENANCE: this capability was advertised in the README and exercised by two validation tests, yet carried NO matrix row and no scenario kind, so it could not be reached from a run at all. It was claimed in prose, absent from the ledger, and unreachable in the engine at the same time. The row and the kind were added together. 0.30 round 2 (pre-registered be137249), blocked (stays MODELLED): ANISE 0.10.6 visibility and a numpy DOP on the row geometry agree (visible sets exact including 72 near-horizon and 48 near-mask cases; DOP within 7.2e-13), but the reported metres are DOP times an illustrative 1.158 m root-sum-square ranging error with no oracle; the scoping that would exclude them awaits an owner decision
Tests
lunar_beacon::tests (horizon visibility against the L01 closed form; the beacon-augmented DOP against the bare one; the error-budget root-sum-square; the DOP-to-metres relation)…
Exact parity-subspace split of a measurement error into the part RAIM cannot see and the part it can. For the linearised snapshot model y = G·x + e, any error dy decomposes uniquely into a BLIND component in range(G) — absorbed as a state error S·dy and annihilated by the residual projector Pperp = I − G·S, so invisible to ANY residual test, not merely to a particular threshold — and a DETECTABLE component in parity space. The module returns the projector, the split, and the blind fraction, so a caller can quantify how much of a specific error a snapshot monitor is structurally unable to report. Applied to the real inter-ephemeris floor: the metre-level DE440-vs-INPOP21a and DE440-vs-EPM2021 disagreement in the geocentric Moon position is absorbed almost entirely as user position error (median blind fraction 1.000000; median blind position error 2.3955 m and 2.0050 m) against a parity residual at the 1e-15 m level
Oracle
P2: numpy linalg.lstsq (LAPACK gelsd SVD, BSD-3) recomputes the least-squares split on committed inputs WITH a real parity part (dy = E·Δs + w, w ~ N(0, 1 m), seed 20261001) on the 8-satellite geometry and a 6-satellite subset, 1464 cases over both real inter-ephemeris pairs, pre-registered (0f63af47): blind_dx, residual, blind_norm and detectable_norm within 1e-12 relative to |dy| and blind_fraction within 1e-12 absolute (worst blind_dx 1.06e-14, residual 4.4e-15, blind_fraction 3.1e-15). numpy builds G itself from the committed positions, while the engine uses normal equations with Gauss-Jordan; an oblique left inverse turns the test red at 5.5e-2 |dy|. Amendment 1 (41ca9bb4, after a first run, disclosed) relaxed only an input precondition on the parity fraction (every case at least 1e-3 and 95 % of each geometry at least 0.05), not the agreement bar. Validated claim, as P2 allows: the linear algebra of the split on those inputs; the inter-ephemeris floor magnitudes and the illustrative lunar geometry (8 LCNS-like nodes at 5000 km) are inputs, not validated, and the blindness is a known property of all snapshot RAIM. The round-1 numpy comparison (tests/lunar_common_mode_integrity_reference.rs) checked only the identity S·G = I because its inputs lay in range(G); it is kept as a regression. Ephemeris provenance: tests/fixtures/inter_ephemeris/NOTICE.md (JPL, IMCCE, IAA RAS)
Tests
lunar_common_mode::tests (8 lib tests: the split is additive and reconstructs dy; a common-mode covariance yields a positive common-mode protection level while a parity-only covariance yields a near-zero one; the…
ModelledCommon-mode protection level and total integrity envelopelunar_common_mode
Capability
cmpl_horizontal bounds the k-sigma horizontal position error contributed by the blind common-mode class, by forming blind_position_covariance = S*Cov*S^T and projecting it into ENU; integrity_envelope reports hpl_total = hpl_araim + cmpl. The distinction the pair exists to make: the RAIM/ARAIM RESIDUAL test is provably blind to range(G) errors, but the ARAIM PROTECTION LEVEL is not - its fault-free term already bounds them up to a provider URA plus nominal bias. The CMPL is therefore only the cross-provider EXCESS that a single provider per-provider URA/ISM does not overbound, and the sum is a conservative triangle-inequality bound rather than an RSS, valid only where that excess is not already inside URA.
Oracle
Analytic projector algebra checked against itself: a covariance confined to range(G) must give a positive CMPL, a covariance confined to parity space must give approximately zero, and the envelope must sum two measurement-space-orthogonal contributions. The common-mode covariance, the constellation geometry and k are representative Modelled inputs, NOT a certified budget, and no claim is made that the triangle bound is tight. NOTE ON PROVENANCE: cmpl_horizontal, blind_position_covariance and integrity_envelope are public and shipped, and their unit tests were already cited by the common-mode blindness row, but the ENVELOPE CLAIM itself carried no row - so the one caveat that keeps it honest, that the sum is a triangle bound and not an RSS, was stated nowhere in the ledger. The row registers the claim and the caveat together.
ValidatedLunar joint multi-technique OD + clocklunar_combination (problem,model,estimate,formal_covariance); batch_ls (gauss_newton_qr)
Capability
Batch fusion of VLBI + lunar-local range + inter-sat range to recover station+constellation positions and clocks
Oracle
P2: numpy 2.4.6 (numpy.linalg.lstsq with LAPACK gelsd, numpy.linalg.inv; BSD-3) recomputes each Gauss-Newton step, the converged weighted-least-squares solution, the formal sigmas and the station error on the committed Jacobians and observations of six simulated networks, pre-registered (8dfbfddd) at max(1e-8, 100 kappa eps) relative. The first run failed the no-VLBI network (6.1e-7 against 1e-8, a normal-equation inverse at condition 1.4e5); the engine fix written after that result (batch_ls::gauss_newton_qr, a Householder QR step used by estimate and formal_covariance) was re-run at the unchanged bar: steps 1.1e-14 to 2.4e-11, sigmas at most 1.3e-11. The validated claim is the fusion solve linear algebra on those inputs, not the physical magnitudes of the simulated network or its observation models: the lunar geodetic VLBI delay row stays MODELLED (its analytic path differs from ANISE by up to 24 µs), and the frames have their own rows. Recovery of an injected simulated truth and NEES consistency remain internal checks; the earlier Orekit 12.2 / Hipparchus Levenberg-Marquardt cross-check (tests/lunar_joint_multi_technique_od_reference.rs) exercises batch_ls::gauss_newton, which estimate no longer calls, and is kept as a check of that primitive. 0.30 revision: the default lunar-joint-od-clock station errors 3.541344 -> 3.541350 m with VLBI and 2183.119 -> 2183.558 m without (QR solve). 0.36 revision (STUDIO-15, portable maths; docs/revisions/STUDIO15-portable-math-rebaseline.md): 3.541350 -> 3.536885 m with VLBI and 2183.558 -> 2184.021 m without, and the no-VLBI solve stops after 5 iterations, not 2 (a round-off-limited stop on the ill-conditioned network); the pre-registered comparison was re-run on re-exported inputs at the unchanged bar (numpy 2.4.6, all six networks pass)
ValidatedFisher information & Cramér–Rao observabilityfim
Capability
Fisher information M=HᵀWH, Cramér–Rao lower bound, observability rank / datum-defect null space (Moore–Penrose pseudo-inverse), and D/A/E/T-optimal experiment-design scalars from a symmetric Jacobi eigensolver
Oracle
NumPy 2.4.1 (numpy.linalg.eigh / .inv; BSD-3-Clause, LAPACK-backed) — an independent third-party authority computing the same uniquely-defined eigenvalues, σ²(XᵀX)⁻¹ Cramér–Rao covariance and GNSS dilution-of-precision factors by a different algorithm (LAPACK divide-and-conquer + LU) than Kshana's cyclic-Jacobi sweep and spectral pseudo-inverse; matched to 1e-9. The same external-library class already validates the DOP engine (vs gnss_lib_py) and the χ²/erf kernels (vs SciPy). Additionally cross-checked against the Kay (1993) closed-form CRLBs and Monte-Carlo CRLB attainment (internal)
Tests
tests/fim_observability_reference.rs (eigenvalues vs numpy.linalg.eigh; CRLB covariance vs σ²(XᵀX)⁻¹ via numpy.linalg.inv; GNSS GDOP/PDOP/HDOP/VDOP/TDOP from the information matrix vs numpy — all matched to 1e-9)…
ModelledLunar absolute-station observability (datum defect)lunar_combination (lunar_observability) + fim
Capability
Fisher-information observability of the joint lunar solve: without Earth baselines the absolute-frame datum lies in the null space of HᵀWH (station position unobservable); ≥3 baselines restore full rank and bound the station Cramér–Rao error, which the estimator attains
Oracle
Analytic datum-defect structure — the unobservable absolute-frame mode lies in the null space of the Fisher information (closed-form), the rank threshold matches the published 3-station design rationale, and the estimator attains the resulting CRLB in Monte-Carlo. The lunar geometry itself is a representative network (not a flown ephemeris), so the row stays MODELLED; the underlying FIM/CRLB engine is checked against the Kay (1993) closed forms separately. 0.30 round 2 (pre-registered e4134aa7), blocked (stays MODELLED): NumPy/SciPy recompute the rank and defect at the stated 1e-9 threshold, the station axes share of the null space (within 8.4e-10) and the station CRLB (within 1.2e-10) on 17 committed Kshana-generated configurations (P2). Not promoted: the claim that the estimator attains the bound (Monte Carlo efficiency about 0.98) has no oracle, and the scoping that would leave it MODELLED awaits an owner decision
Tests
lunar_combination::tests (rank-deficient without Earth baselines; three baselines restore full rank — the 3-station threshold from the information rank, not solve error; station CRLB attained by the estimator at…
7-parameter Helmert datum fit + ICRF orientation tie from a network of points
Oracle
Independent closed-form weighted Umeyama (Horn) similarity-transform solver (numpy/scipy SVD-based; Umeyama 1991 IEEE TPAMI, Horn 1987 JOSA A) — a different algorithm from kshana's iterative Gauss–Newton fit, reaching the same uniquely-defined weighted-LS optimum on byte-identical point networks
Tests
lunar_frame_realise::tests (recovers injected Helmert transform); tests/lunar_reference_frame_realisation_reference.rs (14 cases vs an independent closed-form Umeyama-SVD solver; worst |Δ| translation 2.0e-5 m…
ValidatedLunar navigation service volumelunar_service
Capability
Moonlight-class lunar DOP / coverage / availability + generalised lunar ARAIM protection levels over a service volume
Oracle
Independent third-party authority ANISE 0.10.2 astro::Orbit Keplerian propagator (Nyx Space, MPL-2.0) — an equinoctial two-body code path distinct from kshana's Newton-Raphson Kepler; per-satellite MCI position agrees to <1e-3 m and the derived visible-satellite count/set matches exactly at every grid point. The DOP kernel is gnss_lib_py-validated; integrity uses published LunaNet/LCNS parameters
Tests
lunar_service::tests (DOP reuses the validated kernel; PL reduces to the south-pole case); tests/lunar_navigation_service_volume_reference.rs (per-satellite MCI position to <1e-3 m + EXACT visible-satellite set over…
The Augmented Forward Signal codes of LSIS V1.0 (29 January 2025): the AFS-Q 10230-chip primary generated from the IS-GPS-800 L1C pilot construction with the IS-GPS-800J indices (the same 210 assignments LSIS Appendix D prints), the AFS-I 2046-chip Gold codes generated from g1 = x^11 + x^2 + 1 and g2 = x^11 + x^8 + x^5 + x^2 + 1 with each PRN's G2 delay recovered from the standard's Annex 3 file, the four secondary codes, the 1500-chip tertiary codes, the interim Table 11 node assignment and the tiered pilot code. LSIS content is read at run time from a SHA-256-verified local cache filled from the NASA-hosted PDF, never redistributed
Oracle
Reference: the values LSIS V1.0 prints for validation (Tables C-1 to C-5, D-1 to D-6, E-1 to E-6) and its Annex 3 electronic code files, under a zero-mismatch criterion fixed before the comparison. The criterion is missed by one printed cell that is internally inconsistent with the standard's own Annex 3 file, so the row stays MODELLED; an erratum note is drafted. The same chips also equal LANS-AFS-SIM's (row 'LunaNet AFS baseband rendering against an independent generator, given its channel state')
ValidatedLunaNet AFS baseband rendering against an independent generator, given its channel state (LSIS V1.0)lunar_afs::waveform::render,lunar_afs::codes,lunar_afs::frame
Capability
The rendering core of the AFS baseband (LSIS-130: complex envelope I + jQ, I = the AFS-I Gold code times the 500 symbol/s frame at 1.023 Mchip/s, Q = the tiered AFS-Q pilot at 5.115 Mchip/s, carrier rotation, chip and symbol alignment) reproduces an independent generator's samples when driven by that generator's own recorded channel state (code phases and rates, carrier phase and frequency, per-channel amplitude). Not covered: the node model that derives that state from a Doppler, delay and carrier-to-noise density, the 50/50 power split computed from the carrier-to-noise density (the amplitudes are taken from the simulator), and the SigMF metadata and 8-bit sample writing
Oracle
Library: LANS-AFS-SIM (T. Ebinuma, BSD-2-Clause, commit 480c6bf353717bafc04b5ee5bafb38ed90e61aae, S-band build, run as a separate program); Kshana's samples are driven only by the simulator's recorded channel state, and a print-only trace build is checked byte-identical to the unpatched one. Bars fixed before the run: zero chip and symbol mismatches, |rho| >= 0.999 over the run and >= 0.99 per 2 ms block, phase within 0.05 rad. Equal I and Q amplitudes are an input of the comparison, so the power split is not tested by it. LANS-AFS-SIM reads its tertiary codes from the LSIS Annex 3 file, so that part is not independent of the code row. LANS-AFS-SIM and PocketSDR-AFS share an author and count as one independent implementation. Data-gated: the LANS-AFS-SIM comparison needs the local LSIS V1.0 cache (xval/lunar-afs/fetch_lsis.sh, SHA-256 verified) and skips with a notice without it
Tests
tests/lunar_afs_waveform_lans_afs_sim_oracle.rs::afs_chips_match_lans_afs_sim (the 210 AFS-Q primary codes always; with KSHANA_LANS_RUN and the LSIS cache all 630 codes, zero chip mismatches)…
The 12 s, 6000-symbol FID 0 frame of LSIS V1.0: 68-symbol synchronisation pattern, BCH (51, 8) subframe 1 (octal-763 generator), CRC-24 ((1 + X)·P(X)) on subframes 2-4, rate 1/2 LDPC encoding with the printed A, B^-1, C, D (LSIS Annex 1, read at run time from the SHA-256-verified local cache) and the standard's puncturing, 60 x 98 block interleaver; an exhaustive subframe-1 decoder that flags a TOI outside 0..99, and a normalised min-sum LDPC decoder written from H = [[A, B, 0], [C, D, I]] alone
Oracle
Library: LANS-AFS-SIM (T. Ebinuma, BSD-2-Clause, commit 480c6bf353717bafc04b5ee5bafb38ed90e61aae), whose own BCH, CRC, LDPC and interleaver produce frames symbol-identical to Kshana's for the same content (zero mismatches, bar fixed before the run). Corroborated by Reference values: the LSIS V1.0 Figure 8 BCH worked example and the CRC-24/LTE-A check value (Greg Cook's catalogue of parametrised CRC algorithms), both bit-exact; and by PocketSDR-AFS decoding every subframe of the frames it synchronised in three runs with passing CRC and matching subframe-2 bits. The parity-check test is close to an internal identity (it confirms the printed B^-1 inverts the printed B) and is not counted as an oracle. Data-gated: the LANS-AFS-SIM comparison needs the local LSIS V1.0 cache (xval/lunar-afs/fetch_lsis.sh, SHA-256 verified) and skips with a notice without it
Tests
tests/lunar_afs_waveform_lans_afs_sim_oracle.rs::afs_frames_and_baseband_match_lans_afs_sim (with the LSIS cache: the 3 frames LANS-AFS-SIM transmits, built from the same FID, TOI and data bits: zero of 18 000 symbols…
ModelledLunaNet AFS closed-loop decodability by an independent receiver (LSIS V1.0)lunar_afs::waveform (recording and truth labels),xval/lunar-afs/run_pocketsdr.sh
Capability
A Kshana AFS recording acquired, tracked and decoded by an independent software receiver and compared with the recording's own truth labels. This proves conformance to the standard and label consistency only; it is not a statement about channel physics, propagation, received power or link budgets, whose values in the recording are chosen inputs
Oracle
Library: PocketSDR-AFS (T. Ebinuma after T. Takasu's PocketSDR 0.13, BSD-2-Clause, commit 5b23809f30d68518b7fad7a564fd0fac57cc497d, S-band build, print-only log-level copy), under criteria fixed before each run. The third pre-registered comparison meets its decoding and tracking bars C1 and C2 in full but not its guard G0 as written (two runs of the unmodified receiver differ in the same four acquisition records), and the second missed by 5 violations, so the row stays MODELLED. LANS-AFS-SIM and PocketSDR-AFS share an author and count as one independent implementation; the frame-coding row's printed-vector checks guard against a shared misreading
GPS L1 C/A spreading codes (sdr::CaCode) and LNAV subframes 1-3 (gps_lnav: IS-GPS-200 field quantisation with the specification's pi, two's complement packing, TLM and HOW words, Table 20-XIV parity with D29*/D30* inversion, the parity-solving bits of words 2 and 10) compared bit for bit with gps-sdr-sim on its bundled brdc0010.22n (32 satellites)
Oracle
gps-sdr-sim (MIT, commit 28ca29a6, run as a separate program through a committed harness), pre-registered 45ba9543 with a bit-exact bar on (A) C/A chips, (B) every computed LNAV field, (C) every parity bit. FINDING: (A) 0 of 32 736 chips differ, (C) the parity of 0 of 1 920 words differs, (B) 230 of 1 152 computed fields differ by one unit: gps-sdr-sim's eph2sbf truncates toward zero with C casts where the broadcast integer needs rounding (every differing value lies within 0.01 of a unit of Kshana's rounded integer). Stays MODELLED: the registered bar was the conjunction
ValidatedGNSS IQ scene signal geometry against an independent baseband generatoriq::scene (Scene,SatGeometry,TruthRecord)
Capability
The IQ scene generator's per-satellite truth for a GPS L1 C/A broadcast-ephemeris scene (iq::scene: a broadcast-ephemeris SatGeometry, generated into TruthRecord truth records): which satellites are above the horizon, the pseudorange, the code phase at the first sample, the carrier Doppler and the look angles, for a static receiver, compared with the channel state gps-sdr-sim builds the first block of its own I/Q output from, for the same ephemeris, receiver and start time (ionosphere off on both sides)
Oracle
gps-sdr-sim (T. Ebinuma, MIT, commit 28ca29a6), an independent GPS L1 C/A baseband generator, built with its own Makefile and run as a separate program on its bundled brdc0010.22n at the example location of its usage text (scripts/gen_iq_gpssdrsim_ref.sh); a committed harness linked against the same build prints the channel state (computeRange, computeCodePhase) behind the program's first block. Bars fixed before the generator was run (pre-registered d9ea217d). Scope: GPS L1 C/A geometry for a static receiver without ionosphere, troposphere or receiver clock; the channel models, noise and nav-data content of the scene are not part of this row
Tests
tests/iq_gpssdrsim_cross_generator.rs::scene_geometry_matches_gps_sdr_sim_channels (11 channels: worst pseudorange 1.6e-3 m against 0.05 m, code phase 5.5e-6 chip against 2e-4, Doppler 7.4e-5 Hz against 0.02 Hz, look…
ModelledGNSS IQ acquisition on independently generated I/Q samplesiq::acq (acquire,AcqConfig); iq::io (decode_samples); iq::scene (GpsL1Ca)
Capability
iq::acq::acquire (FFT parallel code-phase search with the square-law Marcum-Q threshold) with the GpsL1Ca replica, on 20 ms of gps-sdr-sim's own 8-bit I/Q output decoded by the iq::io sample decoder (ci8, 2.6 MHz, 11 satellites, no noise): every PRN 1 to 32 searched in two 10 ms windows, the result compared with the code phase and Doppler gps-sdr-sim modulated
Oracle
gps-sdr-sim (MIT, commit 28ca29a6) I/Q samples and its own channel state, pre-registered d9ea217d: (B1) every simulated PRN acquired, (B2) no other PRN acquired, (B3) code phase within 0.5 chip, (B4) Doppler within 125 Hz. FINDING: B1, B3 and B4 hold in both windows (worst 0.195 chip, 72 Hz), but B2 fails: all 21 PRNs gps-sdr-sim did not simulate also cross the threshold (statistic 151 to 206 against 79.5; simulated PRNs 1 300 and up). The threshold is set for Gaussian noise of the measured sample power; gps-sdr-sim adds no noise, so the cells of an absent PRN hold the other satellites' cross-correlation. Stays MODELLED: the registered bar was the conjunction
ModelledIQ monitors scored in lab-replay campaigns: detection, misses, false alarms and closed-form false-alarm figuresiq::campaign (detect,runner,spec); iq::monitor (stats::gaussian_two_sided_pfa,sqm)
Capability
iq::campaign::detect and the campaign runner: an optional [monitors] table runs the IQ monitors on every cell in the runner's single pass. The power and spectral monitors see the samples after each front-end chain; the C/N0, SQM and lock monitors see each tracked channel's loop updates. Each alarm is scored against the stated events: detected or missed per (event, alarm kind, channel), time to detect, and false alarms outside the events over their exposure. Each test's closed-form false-alarm figure is stated: power per block (exact Gamma tail), spectral excess per block (approximation), kurtosis (one-sided Gaussian), pulses (binomial ceiling), SQM per averaged update (2Q(k) = Q(1/2, k^2/2), new), and the C/N0 CUSUM's ARL0 (Siegmund). Without the table, cells, keys and digests are unchanged. No interference waveform is synthesised
Oracle
The monitors' own closed forms on Kshana's own GPS L1 C/A scenes (noise and legitimate signals; attenuation events of the satellites' own signals), against bars pre-registered in docs/design/MONITORS-IN-CAMPAIGN.md (56322c99). FINDING: M1, M2 and M4 met. M3 met on detection (18 of 18 attenuations, P_md 0) and NOT MET on its timing clause: the median time to detect is 1.001 s against 2 x ARL1 samples = 0.433 s. The CUSUM sees one estimate per 1 s window, so the sample-count bound is below one sample. RE-RUN 0.41 (D8b row 8, docs/design/evidence/d8b-iq-timing-campaign-rerun/, 3dcbaace9): with the campaign's automatic C/N0 stride set to one NWPR window (0.020 s), M3 stays NOT MET on its timing clause, as stated before the run: 18 of 18 detected on seen and fresh seeds, median time to detect 0.221 s and 0.201 s against 0.0068 s and 0.0064 s; M1, M2 and M4 unchanged. The miss is structural (the 1 s sliding estimator's delay); a further attempt needs a new pre-registration. v2 (D8b row 8 v2, PREREGISTRATION-v2.md, c46e1479e): with the bound per detected event D_est + 2 x ARL1 x stride, D_est the first stride at which the noise-free drop CUSUM fed the noise-free sliding M2M4 estimate exceeds h, M3 v2 is NOT MET on its timing clause on all three seed sets: 18 of 18 detected on each; median time to detect 0.2208 s, 0.2008 s and 0.2207 s against median bounds 0.0868 s, 0.0864 s and 0.0660 s (seeds 200-205, 100200-100205, 200200-200205; largest bound 0.0916 s, so informative); the independent recomputation of D_est and the bound agrees (worst difference 0 s, bar 1e-9 s); M4 holds (0 false alarms). The cause of the gap between the noise-free and the real estimate is not established. Internal consistency only: no real receiver or recorded interference is checked
Tests
tests/iq_campaign_monitors.rs: m1_power_test_false_alarm_rate_matches_the_closed_form (73 of 5900 blocks against the 99 % binomial interval [40, 80] of P_fa 0.01)…
ModelledDetection ROC of the IQ monitors from a scored lab-replay campaigniq::campaign (detect,report,runner); iq::monitor::spectral (kurtosis_reference)
Capability
iq::campaign::detect (windows, roc) and iq::campaign::report::rocs: each cell of a campaign with monitors keeps, per alarm kind and channel, the maximum of the monitor's own test statistic over each stated event window (positive) and over tiled roc_window_s windows outside the events (negative). The statistics are the block-power departure, the PSD excess, the kurtosis z, the SQM |z| of the averaged metric, and the C/N0 CUSUM. The report pools them by kind into an empirical window-level ROC with its AUC and the operating point at the configured threshold, written as roc.json (kshana.campaign-roc/1) and a Detection ROC section of report.html. A campaign without monitors writes neither
Oracle
Synthetic campaigns from Kshana's own GPS L1 C/A scenes (attenuations of the satellites' own signals), against bars pre-registered in docs/design/DETECTION-ROC.md (04981059). FINDING: R2 met (chance level with nothing to detect). R1 NOT MET: cn0_drop AUC 0.7855 against 0.95, because every negative window after an event inherits the CUSUM's decay from the detection (all 18 post-event negatives reach 81 or more; all 18 pre-event ones stay at 0). RE-RUN 0.41 (D8b row 9, f09660907): R1 MET on re-run with the CUSUM restarted at each scored window's start ([monitors] roc_restart, default on): AUC 1.0000 on seen seeds 300-305 and on fresh seeds 100300-100305; R2 still at chance; the control (restart off) reproduces 0.7855. The seen-seed outcome was largely known in advance (disclosed). Internal consistency only: no real receiver or recorded interference is checked
Tests
tests/iq_campaign_roc.rs: r2_with_nothing_to_detect_the_roc_sits_at_chance (AUC cn0_drop 0.500, sqm_delta 0.566, power_drop 0.639 on stated events that attenuate nothing)…
ModelledCalibration of the open receiver's C/N0 and lock models to IQ campaign results, with stated prediction erroriq::labfit (campaign,cn0,iqfit,iqreport,schema,mod); iq::cli::labfit
Capability
iq::labfit (campaign, cn0, iqfit, iqreport) and [input] campaign on kshana iq labfit, Python kshana.iq_labfit and the MCP labfit tool (campaign folders inside its work directory only): reads a kshana iq campaign output folder by its DIGEST and groups the scored cells into fit units (signal x loop design x front-end chain x jammer type, never pooled). Per unit, it fits measured C/N0 against stated J/S with the spectral-separation form. Q_eff is not identifiable from C/N0 data without an independent measurement of the jammer and signal powers, so Q is stated, never fitted (the type table, the event's q, or for an unknown type the most damaging stated value, cw Q = 1). The fitted parameter is the identifiable offset delta_eff = delta - 10 log10 Q, with the correlation guard reporting the degeneracy. C/N0 bins locked under 90 % are censored and listed. The threshold J/S is given directly and by the closed form; the tracking-loop and empirical lock models are coupled through the C/N0 model, with drop and re-lock dwells. A seeded cluster bootstrap over recordings (1000 replicates) gives percentile intervals; 90 % prediction intervals come from residual draws; leave-one-recording-out hold-outs are split into interpolation and extrapolation; a pre-registered [validation] table is hashed and each bar judged PASS/FAIL. Report kshana.labfit-iq/1 as JSON, CSV and generated Markdown, with a units block. Synthetic only (no third-party recording, no jammer waveform). The receiver-log labfit path is unchanged
Oracle
Synthetic parameter recovery on Kshana's own GPS L1 C/A scenes with a known C/N0 profile imposed from the section 3.1 formula, tracked and scored by kshana iq campaign, against bars pre-registered in docs/design/LABFIT-FROM-IQ.md (section 5.1, amended 2026-10-12 in 4ca3ef3, section 13). The campaigns' C/N0 estimator (M2M4) was fixed in fa388b82, pushed before the first bar evaluation (committed together with the fit code, not alone), after one disclosed diagnostic tracking run. FINDING on the 18-event bar campaign: (c) PASS; (a') NOT MET, delta_eff error -1.097 dB against 0.5 dB (still -0.89 dB on bins at 38 dB-Hz and above); (b') NOT MET, 0 of 50 campaigns cover the truth (error -1.189 +/- 0.093 dB, interval about 0.3 dB); (d) NOT MET, closed form 2.20 dB below the direct J/S at loss, about the fitted drop dwell times the ramp rate (2.09 dB); the original Q clauses of (a) and (b) are NOT MET because Q is not identifiable. FOLLOW-UP (bars pre-registered in 99e7c599, docs/design/evidence/labfit-followup/RESULTS.md): on a stepped-J/S campaign delta_eff is recovered within +0.185 dB (S1 MET: the ramp error is estimator lag); (d) as written on it is met by the tracking-loop model (0.45 dB) but not the empirical one (6.42 dB), so S2 is NOT MET; on a fresh ramp campaign (d) with the fitted drop dwell times each event's ramp rate added is within 1 dB for both models (R1 MET), and (a') and (d) as written miss again (-1.25 dB; 2.32 and 9.47 dB). RE-RUN 0.41 (D8a row 6, 61b389998, docs/design/evidence/d8a-iq-tracking-cn0-rerun/): with the fit modelling each bin 0.5 s earlier along the profile ([campaign_fit] cn0_lag_s), still NOT MET: (a') +0.5015 dB seen (bar 0.5) and +0.399 dB fresh; (b') 0 of 50 on seen and fresh seeds, now +0.45 to +0.47 dB with every interval above the truth (the 0.5 s lag overshoots); (c) met on seen seeds, missed on fresh (time-to-loss coverage 1.000 against a bar of at most 0.98); (d) 2.19 dB seen, 0.32 dB fresh. The controls (lag 0) reproduced before scoring (disclosed). A different lag would need a new pre-registration. RE-RUN v2 (PREREGISTRATION-v2.md, scored once, 7892f3a24; determinism re-check of batch 17 on a second machine byte-identical before scoring): the fit models the estimator's own response, with no free parameter; still NOT MET: (a') met on all three sets (+0.165, +0.098, +0.099 dB); (b') 36, 32 and 27 of 50 against a bar of 45 (errors +0.083 to +0.099 dB against intervals about 0.23 dB wide); (c) time-to-loss coverage 1.000 on the v1-fresh set (bar at most 0.98), met on the other two; (d) 2.195 dB on the v1-registered set (bar 1 dB), 0.324 and 0.193 dB on the fresh sets. The method is checked, not any receiver: it cannot be VALIDATED from synthetic data
Tests
tests/iq_labfit_campaign.rs: bar_c_holdout_prediction_intervals_cover_between_80_and_98_percent (PI90 coverage 0.891 over 558 bins, 0.900 over 10 time-to-loss events under each lock model)…
ModelledGPS LNAV encoding decoded by an independent receiver librarygps_lnav
Capability
Kshana's LNAV subframes 1-3 (gps_lnav::encode_subframes) for the 32 brdc0010.22n ephemerides fed to RTKLIB's decode_word and decode_frame: parity of every word, subframe decoding with consistent issue of data, and every decoded parameter against Kshana's integer and against the input value
Oracle
RTKLIB v2.4.2-p13 (BSD-2-Clause, commit 71db0ffa) decode_word/decode_frame as a separate program, pre-registered c26e2b28: R1 parity and decoding, R2 decoded value equal to Kshana's integer times the IS-GPS-200 scale within 2^-52 relative, R3 within half a quantum of the input. FINDING: R1 960 of 960 words and 96 of 96 subframes; R3 every parameter, week and reference times exact; R2 fails on 41 values (af1, delta-n, IDOT) by 1 to 4 units in the last place because RTKLIB's P2_43 is the decimal 1.136868377216160E-13, 2.6e-16 relative from 2^-43; dividing by RTKLIB's constant returns Kshana's integer in every case. Stays MODELLED: the registered R2 bar was too tight for the oracle's own constant
IS-GPS-200 LNAV subframes 1-3 (gps_lnav::encode_subframes: round-to-nearest quantisation of every clock and orbit parameter at the specification's scale factors and value of pi, two's complement packing, TLM and HOW words, Table 20-XIV parity with D29*/D30* inversion, the parity-solving bits of words 2 and 10), the message a GPS L1 C/A signal generator transmits
Oracle
RTKLIB v2.4.2-p13 (BSD-2-Clause, commit 71db0ffa) readrnx parsing the IGS daily broadcast file BRDC00IGS_R_20250610000_01D_MN.rnx and decode_word/decode_frame decoding Kshana's words, run as a separate program; pre-registered a48d5b91 before the file was fetched, exact-integer bar. Mutation: truncating instead of rounding gives 235 failures. Earlier comparisons of the same encoder are recorded as findings: against gps-sdr-sim, which truncates (230 fields), and against RTKLIB's scaled values, whose decimal 2^-43 constant missed a one-ulp bar (41 values). Integrator post-verification (stays MODELLED): the comparison agrees and a parity-set mutation also turns it red, but the pre-registration a48d5b91 reached the public branch at 20:02:29 UTC, two seconds after the result commit 25638e9e (20:02:27 UTC) that already holds RTKLIB's output for the file, so the order 'registered before the file was fetched' cannot be shown; a fresh registration pushed before an unseen broadcast file is fetched would settle it
Tests
tests/gps_lnav_rtklib_integer_oracle.rs::kshana_lnav_integers_decode_in_rtklib_from_an_unseen_broadcast_file (32 satellites of the IGS broadcast file of 2 March 2025: 960 of 960 words pass RTKLIB parity, 96 of 96…
ModelledAcquisition and C/N0 estimation on real lunar IQacquisition,portable_math,realdata::ion_sdr,realdata::lugre
Capability
Parallel code-phase search (acquisition::pcps_acquire: portable FFT, coherent folding, non-coherent accumulation, chi-square threshold for a search-wide false-alarm probability), refinement, open-loop prompt correlation with code Doppler and the M2M4 C/N0 estimator, run on the LuGRE L1 4-bit IQ batches read through realdata::ion_sdr
Oracle
GNSS-SDR 0.0.19 (GPL-3.0, separate program, pinned configuration) on the nine non-surface L1 batches and the flight receiver's RAW C/N0 (LuGRE Mission Data, Zenodo 16411687, CC BY 4.0), pre-registered 7425dc34. FINDING: Kshana's reader decodes all 288 compared spans identically to an independent converter; GNSS-SDR declares 266 of 288 searches positive and Kshana agrees on all 266 decisions, but only 250 are located on GNSS-SDR's cell (A2 fails); the C/N0 leg is vacuous, no flight RAW epoch lying within 30 s of any batch (nearest 141 s). Most positives are artefacts: the bare two's-complement levels carry a -0.5 mean on I and Q, a zero-frequency line 32 dB above the noise in a 1 kHz bin, which both receivers acquire. Two batches (OP5, OP12) were excluded in error: a digit-stripped reading of their metadata was taken for a contradiction with their headers; the real cause was that the reader could not read a sample wider than a word (since fixed). Stays MODELLED
ModelledAcquisition on real lunar IQ against orbit-predicted Doppler (sample-power decision)acquisition,realdata::ion_sdr
Capability
acquisition::pcps_acquire with its sample-power decision on the nine non-surface LuGRE L1 batches with mid-rise levels; Doppler differences between acquired satellites compared with differences predicted from ESA final orbits and the Firefly-reconstructed Blue Ghost trajectory
Oracle
Predicted Doppler from ESA/ESOC final orbits and the NAIF CLPS reconstructed trajectory (xval/lugre-predicted-doppler/predict.py, independent of Kshana), pre-registered c9868b3e. FINDING: all 288 searches cross the threshold on noise (sample-power statistics about 430 to 520 against 346.3), so P1 holds trivially and P3 holds (274 of 288 predicted visible) while P2 fails on 4033 of 4036 pairs. The band-limited front end raises every cell's noise about 1.46 times (pinned: noise floor 292 against the white-noise 200 on OP2 PRN 1), which the sample-power statistic does not see. Stays MODELLED
ModelledAcquisition on real lunar IQ scored by orbit-predicted Doppleracquisition,realdata::ion_sdr
Capability
acquisition::pcps_acquire with the cell-averaging decision (peak over the grid's mean cell, robust to band-limited noise) on the nine non-surface LuGRE L1 batches with mid-rise levels; Doppler differences between acquired satellites compared with differences predicted from ESA final orbits and the Firefly-reconstructed Blue Ghost trajectory
Oracle
Predicted Doppler from ESA/ESOC final orbits and the NAIF CLPS reconstructed trajectory (xval/lugre-predicted-doppler/predict.py, independent of Kshana), pre-registered 470b0fc9. FINDING: P1 and P3 hold (4 batches, 12 of 12 acquisitions predicted visible) but P2 fails on 8 of 8 pairs. The two strongest pairs match the prediction in magnitude with the sign reversed (OP2 PRN 18/24: +7670.0 Hz measured, -7693.9 predicted; OP17 PRN 11/30: +11155.0, -11114.4), so the registered I-in-the-low-nibble order is the conjugate of the batches; the six detections just above threshold match neither sign. The refinement's Doppler scatter (30 to 50 Hz between two runs) was also underestimated by the registration. Stays MODELLED
ModelledAcquisition on held-out real lunar IQ against orbit-predicted Doppleracquisition,realdata::ion_sdr,realdata::lugre
Capability
acquisition::pcps_acquire (cell-averaging decision, false-alarm probability 1e-7) with refine_doppler_coherent on the LuGRE L1 batches recorded on the lunar surface (I in the high nibble, mid-rise levels); Doppler differences between acquired satellites against differences predicted from ESA final orbits and the Firefly landing-site kernel
Oracle
Predicted Doppler from ESA/ESOC final orbits and the NAIF CLPS Firefly landing-site kernel (xval/lugre-predicted-doppler/predict_surface.py, independent of Kshana), pre-registered 73ee3e14 on batches never opened before. FINDING: non-vacuity fails; no batch holds two acquisitions. OP40 excluded (metadata names a missing file); OP38, OP73, OP77_0, OP77_1 acquire nothing and show no false alarm at 1e-7 (top statistics 325 to 356 against 385.9); OP74 and OP76 acquire PRN 31 and OP78_0 and OP78_1 PRN 12, each alone, all predicted visible. Stays MODELLED. Disclosed deviations: the registered 10 Hz Doppler bar rests on a synthetic Monte Carlo of the coherent refinement, not on the scatter of the already-seen batches the integrator asked for; OP40 was excluded before the run because its metadata names a sample file the dataset does not ship
Per-satellite topocentric look angles (azimuth clockwise from north, elevation) and slant range at a named selenographic site, given the satellite in the site's Moon body-fixed frame (the validated claim). The signal-in-space ranging accuracy is exposed as a scenario parameter so the service-volume sweep yields a ranging REQUIREMENT rather than a pass/fail at one fixed sigma; that parameter is a design input outside the validated claim
Oracle
ANISE 0.10.2 (Nyx Space, MPL-2.0, run as a tool from xval/anise-service-geometry, binary lunar-look-angles-xval, which calls no Kshana code): ANISE's own azimuth, elevation and range routine (azimuth_elevation_range_sez) from sites built by its latitude-longitude-altitude constructor on the pck00011 Moon (sphere 1737.4 km), with the satellites rotated into MOON_PA_DE440 through moon_pa_de440_200625.bpc. 768 samples over 8 sites (the 4 named sites and 4 others, above and below the horizon), 12 epochs and 8 satellites agree within 1e-6 deg in azimuth and elevation and 1 mm in range and site position (fixed before the comparison); measured 2.2e-12 deg, 2.1e-13 deg and 5.6e-9 m. Claim: site placement and look angles given the satellite in the site's body-fixed frame. Kshana's own mean-rotation Moon-fixed frame realisation and the ranging-accuracy parameter are outside it; the closed-form geometry unit tests remain as supporting checks
Tests
lunar_service::tests (topocentric against hand-computed geometry: overhead, due north/east/west, antipodal, and the degenerate polar east direction; the exported visible flag agrees with the independent visibility…
Differential error-cancellation identity + reuse of the DO-229E SBAS PL machinery; the single-difference + WLS solve is additionally cross-checked against RTKLIB's lsq()/matinv() (Takasu, BSD-2-Clause, compiled C) — independent solver code, but the same first-order LOS-difference algebra, so still InternalConsistency
Tests
lunar_dpnt::tests (clock common-mode cancels exactly; residual grows with baseline; reuses SBAS PL; the satellite count is honoured to the builder's limit of 24, so a larger constellation cannot silently return a…
LunaNet/IOAG-aligned lunar frame + time + ephemeris export (CCSDS OEM + KIF) with round-trip conformance
Oracle
kshana's lunar OEM export re-parsed by the independent third-party oem library (R. J. Anderson): frame/time tokens and per-epoch state agree to write precision (1 mm / 1e-9 km/s) and a dropped-TIME_SYSTEM export is rejected — a structural interchange round-trip; the lunar frame/time physical semantics are validated by their own rows, so this stays MODELLED. 0.30 external comparison, a finding (stays MODELLED): a second independent reader, the Orekit 12.2 CCSDS OEM parser (Apache-2.0, strict defaults), refuses both lunar files because it does not implement the LTC and TCL time systems (it decodes the crate's EME2000/UTC OEM, and decodes the lunar files once TIME_SYSTEM is replaced by TDB); oem 0.4.5 decodes them within the written precision only by falling back from the unsupported time system. Interchange of these lunar files with mainstream CCSDS readers is not established
Tests
lunar_interop::tests (OEM carries lunar REF_FRAME/TIME_SYSTEM; time metadata round-trips; KIF envelope); tests/lunar_interoperability_export_reference.rs (kshana's emitted lunar OEM re-parsed by the independent oem…
Per-dimension sub-scores over Kshana's own seven resilience categories (not DHS terms), RethinkPNT RDRR functions and Yang criteria, each tagged Modelled with its driver; tentative RPCF Level with a bounded-degradation gate. Simulation-derived self-assessment, never certification.
Quantifies how a single composite score / RPCF Level reorders architectures under a defensible weighting simplex and a threat ensemble (top-1 flip rate, Kendall-tau dispersion, Level-flip rate, rank ranges); declared-vs-measured and diversity-collapse analyses.
Oracle
Closed-form rank-statistics identities (tau in [-1,1] with hand-computed values; deterministic seeded Dirichlet) and constructed stable/unstable witnesses
Tests
resilience::stats::tests (Kendall-tau hand example, Dirichlet simplex, flip-rate); resilience::study::tests (stability control, instability witness, declared-vs-measured, diversity collapse)
Per-result honesty ledger qualifying a demonstration output: its external anchors, modelled assumptions, gaps-to-flight and representative TRL band, with invariants enforced (Validated requires an external anchor; Modelled requires a gap and cannot claim above TRL 4).
Oracle
Closed-form invariants mapping to the 'representativeness justified + gaps-to-flight identified' compliance discipline; tied to the verification status/oracle-kind boundary
One reproducible TradeEvidence object (fixed frame: scenario+seed+engine; common per-FoM quantum-vs-classical values with polarity-correct benefit, optional 95% CI, validated/modelled label) carrying a representativeness record, so every quantum-PNT vertical reports the trade the same honest way.
Oracle
Closed-form benefit/winner identities + faithful wrap of the existing quantum_trade::TradeResult; honesty tied to the representativeness ledger and verification labels
Tests
qtrade::tests (benefit polarity higher/lower-is-better, wraps a real TradeResult faithfully, dishonest evidence rejected, validated-FoM needs external anchor, deterministic JSON)
Device cards (optical/trapped-ion/mercury-ion + classical clocks reused from holdover/clock_state; cold-atom interferometer; classical + entanglement/single-photon time-transfer links) each carrying a representativeness record; the entanglement link adds a shot-limited timing-precision model (~jitter/sqrt(R*tau), dark-count penalty, systematic floor).
Oracle
Reused clock/CAI coefficients (holdover/clock_state, published values) + closed-form shot-noise/loss identities for the entanglement link
ModelledTrusted quantum timing (time transfer + secure dissemination + anomaly)timetransfer_chain
Capability
End-to-end quantum vs classical time-transfer chain (clock coast + link precision in quadrature), a reused timing protection level, a delay/replay-attack security FoM (1-P_md) and a clock-anomaly detection probability + CUSUM latency, emitted as honest TradeEvidence with a representativeness record.
Oracle
Closed-form quadrature budget over reused validated kernels (ADEV vs Stable32/NIST; TPL bound; detection analytic_pd/pmd); honesty tied to the representativeness ledger
Tests
timetransfer_chain::tests (precision improves with integration; quantum can win AND lose; PL finite-positive; security FoM in [0,1] and grows with attack delay; anomaly Pd monotone; trade is_honest)
Quantum (cold-atom interferometer) vs classical navigation-grade INS dead-reckoning over a GNSS outage: position-error growth, holdover to a position threshold, and the quantum-vs-classical trade as honest TradeEvidence; honest observability note (bias unobservable without a fix, so error grows).
Oracle
Reused inertial budgets cross-checked against an independent Octave double-integration of the same dead-reckoning ODE (different runtime, shared model → ReferenceImpl) plus the Freier-2016 published short-term noise as a one-sided anchor; the quantum-vs-classical composite stays MODELLED
Tests
quantum_nav_od::tests (quantum beats classical over a long outage; advantage is outage-dependent; trade is_honest); tests/gnss_free_quantum_navigation_reference.rs (dead-reckoning position growth vs an independent…
ModelledFault/anomaly detection for quantum PNT systemsquantum_faults
Capability
Labelled quantum-fault catalog (clock frequency-jump/drift/lock-loss; sensor bias-step/dropout), a detection-statistic ROC AUC with a bootstrap CI, and a minimum-detectable fault at a fixed false-alarm rate; a quantum-clock-aided monitor detects smaller faults than a classical one, emitted as honest TradeEvidence.
Oracle
Closed-form Gaussian AUC = Phi(mu/(sigma*sqrt2)) cross-checked against the externally-validated eval_stats::bootstrap_auc_ci (vs scikit-learn) + detection analytic thresholds
Tests
quantum_faults::tests (analytic AUC known values; empirical bootstrap AUC brackets the closed form; quantum detects smaller faults / higher AUC; advantage vanishes for huge faults; 5-class catalog; trade is_honest)
Euler's rotational equations of motion (I ω̇ = τ − ω × Iω, principal-axis and general inertia tensor) coupled to quaternion attitude kinematics (q̇ = ½ q ⊗ ω) and propagated with a fixed-step RK4 integrator that re-normalises the quaternion each step
Oracle
Basilisk 2.9.1 spacecraft hub rigid-body propagation (AVSLab, ISC; MRP attitude through its own coupled hub equations, svIntegratorRKF78 at relative tolerance 1e-12), an independent implementation of the torque-free rigid-body equations: Kshana's RK4 quaternion (dt 0.01 s) and body rates agree within 1e-9 over 1e4 s on two general non-diagonal inertias (worst 3.3e-11 and 3.4e-12 rad/s); tolerance fixed before the comparison. Validated claim: torque-free motion; the external-torque term is not compared. The conservation-law and symmetric-top checks remain as supporting self-consistency. No flexible-body / control-loop / external-torque environment
Tests
attitude_dynamics::tests (apply/solve inverse, spherical-top zero torque, principal-axis fixed point, short-run energy+momentum conservation, q̇=½q⊗ω, symmetric-top rate sign + body-cone precession)…
ValidatedClohessy–Wiltshire / Hill relative-motion dynamicscw_dynamics (propagate,second_order_correction,propagate_second_order)
Capability
Relative motion of a chaser about a target on a circular reference orbit in the LVLH frame: the linearised Clohessy–Wiltshire equations (ẍ−2nẏ−3n²x=0, ÿ+2nẋ=0, z̈+n²z=0) solved by the closed-form 6×6 state-transition matrix Φ(n,t) (Clohessy–Wiltshire 1960; Vallado Alg. 48) with the bounded relative-orbit condition ẏ₀=−2n·x₀, and a closed-form second-order correction (cw_dynamics::second_order_correction, the CW response to the quadratic terms of the exact circular-chief equations; Karlgaard and Lutze 2003, Newman, Lovell and Pratt 2015) composed as cw_dynamics::propagate_second_order. The validated output is propagate_second_order; the linear Φ alone is characterised against nonlinear truth, not validated
Oracle
Orekit 12.2 (Apache-2.0) NumericalPropagator, two-body, LOFType.QSW, run as a separate program (fixture unchanged since round 1). Pre-registered bar 1e-3 m on every sample of every case; the second-order engine fix was re-run against it unchanged (amendment dec81b7c, written before the propagator existed, disclosing that the first-order gaps were known). propagate_second_order agrees within 1.15e-7 m (C1), 9.8e-8, 5.4e-8 and 2.2e-8 m; halving the second-order term fails at 1.85e-3 to 3.14e-3 m. The linear Φ alone departs from nonlinear truth by 1.2 to 6.3 mm at 100 m over a third of an orbit (the omitted second-order term; its strict test is kept ignored), so it is characterised, not validated. Internal checks kept: Φ against an RK4 integration of the Hill equations, Φ(t)Φ(−t)=I, the bounded-orbit and −12π·x₀ drift identities, and the third-order residual scaling of the second-order term. Scope: circular chief, separations near 100 m, up to a third of an orbit; no eccentricity (Tschauner–Hempel), J2 or differential-drag terms; small-separation assumption
Tests
cw_dynamics::tests (Φ(0)=I, cross-track decoupled SHM); tests/cw_dynamics_reference.rs (closed-form Φ vs an independent fixed-step RK4 integration of the same Hill ODEs to <1e-6 over a third of an orbit; Φ(t)Φ(−t)=I to…
Locate an emitter (jammer/spoofer, or an opportunistic source for reverse-PNT) from time-difference-of-arrival across a receiver network — the τᵢ=(Rᵢ−R₀)/c hyperboloid intersection solved by Gauss–Newton least squares — and, adding frequency-difference-of-arrival (range-rate differences) with moving receivers, jointly recover position and velocity; with the Cramér–Rao lower bound on the position covariance from the measurement geometry
Oracle
Self-consistency of the estimator and geometry: forward→inverse round trips, the Fisher/CRLB identity J·CRLB=I, GDOP monotonicity, and the estimator attaining its own Cramér–Rao bound under Monte-Carlo noise — internal-consistency checks, NOT an external dataset, so the row stays InternalConsistency. MODELLED passive geolocation — point-source line-of-sight model; no multipath / NLOS, receiver-clock-bias, or atmospheric-refraction terms. 0.30 round 2, against Ho and Chan 1993 and Ho and Xu 2004 (pre-registered 034e1d1b, 1 % of each printed value), a finding (stays MODELLED): the surface-constrained TDOA bound reproduces the 40 N geometric factors (-0.11 %, -0.02 %) and the maximum TDOA; the joint TDOA+FDOA bound reproduces Fig. 6 (+0.06 %, -0.17 %) and, at the source the authors own code names for Fig. 7 (amendment written after that result was seen, disclosed), Fig. 7 (+0.11 %, -0.02 %); it equals the authors published bound code (BSD-style, run under GNU Octave) to 3e-12; the covariance-weighted maximum-likelihood solver attains the printed bound (+0.3 to +0.4 %). The Ottawa factors the paper prints from a Monte Carlo (251, 7.1) differ by +1.3 % and +4.4 %, outside the 1 % bar
Tests
geolocation::tests (noiseless TDOA forward→inverse to 1e-6 m; J·CRLB=I with a symmetric PD covariance; the CRLB position-variance trace is non-increasing when a receiver is added; joint TDOA+FDOA recovers a moving…
Validated3-DoF ballistic entry peak deceleration against the published Stardust nominalentry_eo_ascent::entry_heating (simulate_entry,inertial_to_relative,EntryHeatingScenario)
Capability
A 3-DoF point-mass entry over a rotating spherical Earth through the US Standard Atmosphere 1976, integrated with RK4 from an inertial entry state, reporting the peak deceleration and its altitude; the planet-relative entry state from the inertial one. Runnable as the entry-heating scenario kind.
Oracle
The Stardust nominal of Desai, Mitcheltree and Cheatwood (1999, NASA NTRS 20050042027): maximum deceleration 33.4 g from 12.6 km/s planet-relative with the 45.8 kg launch mass, a 6-DoF POST trajectory with the full aerodynamic database; entry interface from Desai, Lyons, Tooley and Kangas (NTRS 20090004445): radius 6503.14 km, inertial 12.9 km/s at -8.2 deg. Bars pre-registered in xval/entry-stardust/PREREGISTRATION.md before any run on these inputs; both held on the first run. WHAT IS NOT: one published case and one quantity (peak deceleration, which in the ballistic approximation does not depend on the ballistic coefficient); the entry latitude and heading are assumed; heating, heat load and the landing point are not checked against it; point mass with constant C_D, not 6-DoF.
Tests
tests/entry_stardust_published.rs (peak 35.44 g against the published 33.4 g, +6.1 %, bar 10 %; planet-relative entry speed 12.551 km/s against 12.6, bar 0.1 km/s); entry_eo_ascent::entry_heating::tests (rotating-frame…
ModelledStagnation-point convective entry heating (Sutton-Graves) on a 3-DoF entryentry_eo_ascent::entry_heating (stagnation_heat_flux,simulate_entry)
Capability
Stagnation-point convective heat flux q = k sqrt(rho/R_n) V^3 with the Sutton-Graves constants for Earth air and the Mars atmosphere (NASA TR R-376, 1971), its peak and the integrated heat load along a 3-DoF entry, ballistic or lifting at constant bank, on Earth or Mars.
Oracle
Internal consistency: the relation and constants as printed in NASA TR R-376, a half-step convergence check, and the expected ordering of heating and deceleration peaks. WHAT IS NOT: no published heating value has been compared (the Stardust papers give a design limit and a convective-plus-radiative estimate of about 1200 W/cm2, which a convective-only relation cannot match); no radiative heating, wall-temperature or catalycity correction, or thermal-protection response.
Nadir imager signal-to-noise budget (Lambertian scene radiance, camera equation, shot, dark, read and quantisation noise), the system MTF at Nyquist as the product of unobscured-aperture diffraction, detector aperture, linear smear and Gaussian jitter, the edge response and relative edge response integrated from that MTF, Q = lambda F / p, and the GIQE version 4 NIIRS with no MTF compensation. Runnable as the eo-radiometry scenario kind.
Oracle
Closed forms: the diffraction MTF at half cutoff, the one-pixel ramp edge of a pure detector aperture (RER 1), the normal-CDF edge of a Gaussian blur (RER = erf(0.5/(sigma sqrt 2))), shot-noise scaling, and the GIQE 4 coefficients as published (Leachtenauer et al., Applied Optics 36(32), 1997). WHAT IS NOT: no external imager, measured MTF or rated image has been compared; the GIQE is an empirical regression with its own error; no stray light, atmospheric MTF, MTF compensation or aliasing.
ModelledLaunch ascent gravity, drag, steering and back-pressure lossesentry_eo_ascent::ascent (simulate_ascent,AscentScenario); reentry::us76_pressure
Capability
Planar gravity-turn ascent of a multi-stage vehicle over a non-rotating spherical Earth through the US Standard Atmosphere 1976 (density and pressure), with a pitch kick, thrust reduced by the nozzle back-pressure, stage burnouts at exact step boundaries, and the gravity, drag, steering and back-pressure losses integrated beside the state. Runnable as the ascent-losses scenario kind.
Oracle
Internal consistency: the exact identity burnout speed = sum Isp g0 ln(m0/mf) minus the four integrated losses (residual below 1e-6 m/s), Tsiolkovsky's equation on a vertical vacuum burn with gravity the only loss, and the US76 sea-level and tropopause pressures. WHAT IS NOT: no published vehicle's losses have been compared; no Earth rotation, coast, throttling or guidance beyond the kick; constant C_D; not a trajectory optimiser.
ModelledTWSTFT two-way equation, uncertainty budget and link stability with flicker and diurnal non-reciprocitytwstft_budget (counter_readings,solve_offset,combined_uncertainty,flicker_phase,two_way_series,series_stability); time_transfer_budget (TimeTransferBudgetScenario,read_series)
Capability
The TWSTFT two-way equation solved for T_A - T_B with station transmit and receive delays, the Ku-band ionospheric up/down asymmetry, the transponder delay difference and the Sagnac terms, against its exact forward model; a quadrature uncertainty budget with the link TDEV as the type A term; a two-way series with white and flicker phase noise (Kasdin-Walter), a diurnal non-reciprocity and a constant reciprocity residual; and the rate, residual RMS, TDEV and ADEV of supplied Circular T, MJD-CSV or IGS RINEX station-clock series. Runnable as the time-transfer-budget scenario kind.
Oracle
Internal consistency: the forward model and the solved equation agree to 1e-15 s, the Sagnac term equals the module's independent BIPM 2A omega/c^2 loop form, the ionospheric term is 40.3 TEC (1/f_up^2 - 1/f_down^2)/c, and the noise processes show their expected TDEV shapes. WHAT IS NOT: no real TWSTFT measurement or calibration report has been compared; the stated delays and uncertainties are inputs, not measured; the stabilities of supplied series are reported, not validated.
Tests
twstft_budget::tests (the two-way equation recovers the offset exactly; the Sagnac term equals the BIPM loop closed form; the ionospheric term follows the frequency split and the troposphere cancels; flicker phase has…
Hufnagel-Valley Cn2 integrated along a slant path for the plane- and spherical-wave Rytov variance (uplink and downlink), the Andrews-Phillips scintillation index and gamma-gamma parameters, the gamma-gamma fade probability and fade margin, the weak-turbulence aperture-averaging factor, the cloud optical depth of a liquid-water column and the Kim-visibility plus Rayleigh clear-air optical depth, replacing the lumped atmospheric loss of the optical link budget when an atmosphere is supplied. Runnable inside the optical-atmosphere scenario kind.
Oracle
The numerics are compared with an independent program (xval/optical-atmosphere/oracle.py, mpmath 1.4.1 at 30 digits, tanh-sinh quadrature on the original height variable and Bessel-K gamma-gamma densities) that shares no code with Kshana, under bars fixed before any implementation existed (xval/optical-atmosphere/PREREGISTRATION.md). WHY MODELLED: that oracle evaluates the same published formulas, so it checks the implementation of the stated formulas and not the atmosphere; no ITU-R validation example and no measured scintillation or slant-path extinction record is used (the ITU values are an open licence decision and are kept out of the repository). The Cn2 profile level, the visibility, the aerosol and molecular scale heights, the aperture-averaging path length and the moment-matching used to apply aperture averaging to the gamma-gamma parameters are stated inputs or conventions. A source line in the module records that one widely copied survey prints the second scintillation exponent as 7/6 where Andrews and Phillips give 5/6; the 5/6 form is implemented
Tests
tests/optical_atmosphere_reference.rs (pre-registered bars A1-A3 against the independent mpmath oracle: 216 Rytov grid points worst relative error 3.4e-7 against a 1e-6 bar, scintillation worst 4.5e-16, aperture factor…
ModelledWeather-driven optical ground-site availability and distance-correlated site diversityoptical_weather; optical_sitediv; optical_techniques; optical_atmosphere_scenario
Capability
Availability of an optical link against a stated link margin from a log-normal cloud liquid-water climate (ITU-R P.840 amount, geometric-optics extinction) and the independent gamma-gamma turbulence fade, the link margin needed for target availabilities and the ceiling no margin lifts, log-normal outage durations, enhancement-technique what-ifs, and a one-factor Gaussian-copula network availability whose pair correlation is exp(-distance/decorrelation length). Runnable as the optical-atmosphere scenario kind.
Oracle
Internal consistency: closed-form limits and monotonicity of the diversity model, the cloud-ceiling closed form, and the log-normal outage identities. NO external oracle is claimed and none has been compared: no measured optical-ground-station availability or cloud-free-line-of-sight record is used. WHY MODELLED: the cloud climate (probability of cloud, median and spread of the liquid-water column, outage durations), the independence of cloud and turbulence, the Gaussian-copula form, the decorrelation length and every technique gain are assumptions; the bundled scenarios state ILLUSTRATIVE climates, not a dataset. Upgrading this row needs a cited measured availability record and a bar registered before the comparison
Tests
tests/optical_atmosphere_identities.rs (bar A4: rho = 0 equals the independent union and rho = 1 the best single site to 1e-12, and zero monotonicity violations in the number of sites and in the correlation; the first…
ModelledTiming and ranging continuity through optical outagesoptical_outage_holdover (couple,clock_psds); holdover (coast_phase_sigma,holdover_seconds,phase_to_range_m)
Capability
The outage-duration distribution of an optical site coupled to the closed-form coast variance of the clock holdover model: timing and ranging error at stated outage quantiles, the holdover to a timing threshold, the fraction of outages and of outage time inside it and the timing availability A + U E[min(D, T_h)]/E[D], for the classical and quantum clock classes. Runnable inside the optical-atmosphere scenario kind.
Oracle
Internal consistency with the holdover module, bit for bit; NO external oracle is claimed. The class PSDs are the synthesised long-tau floors the holdover module already carries, so the figures inherit its stated caveat that holdover to a tight threshold is governed by the assumed red-noise floor and not by the cited Allan deviation. The outage durations are a log-normal model with stated parameters, not a measured cloud-dwell record
Tests
tests/optical_atmosphere_identities.rs (bar A6: the per-outage timing error equals holdover::coast_phase_sigma and the range error equals phase_to_range_m bit for bit over 9 clocks, 3 mean durations, 3 spreads and 4…
ModelledOptical link margin against time: one seeded realisation of a site's weatheroptical_timeseries (site_series,SiteModel,SeriesSpec,Series); optical_atmosphere_scenario (TimeseriesSpec)
Capability
An opt-in [timeseries] table (duration, step, seed) makes the optical-atmosphere kind also write, per site, one seeded realisation of the stated site model through time: cloud episodes (clear spells exponential, cloud outages log-normal with the stated mean and spread, the share of time in cloud set to the stated cloud probability), one log-normal liquid-water column per cloud episode, one gamma-gamma scintillation sample per step, and from them the link margin, the cloud and scintillation fades and the link-up flag, with the realised availability, cloud fraction, outage count and longest outage. Without the table the result is the 0.37 document. Draws go through the portable distribution and maths routes, so native and wasm32 give the same bytes.
Oracle
Internal consistency only; NO external oracle is claimed and none exists for a weather realisation. The checks compare one draw with the statistical output of the same model, so a pass says the generator samples the stated distributions (the draw's availability and cloud share sit inside pre-registered tolerances of the stated figures), not that any real site behaves so. The series is one realisation, not the climate: its availability differs from the stated one by the draw's sampling error, and scintillation samples at different steps are independent, a stated simplification of a process that decorrelates in milliseconds. The weather inputs stay the scenario's assumptions.
Tests
tests/optical_atmosphere_timeseries.rs (bars T1-T6 of xval/optical-atmosphere/PREREGISTRATION-timeseries.md, fixed before the first run: same seed same bytes and another seed other bytes; no key without the table; over…
ModelledTRK-2-18 Orbit Data File reader and Mars ramped 2-way Doppler modelodf (parse_odf,Odf,OdfObservation,OdfRamp,seconds_1950); mars_od (DopplerModel,RampTable,compress_two_way,MarsDynamics,fit,fit_damped); naif_kernel (SPK type 1); gravity_pds (SphericalHarmonicField::from_shadr)
Capability
A reader for DSN Orbit Data Files (TRK-2-18): the 36-byte record groups (file label, identifier, orbit data, per-station ramps, clock offsets, summary, end of file), and in each orbit-data record the time tag since 1950, downlink and uplink delays, observable, stations, data type, bands, validity, channel, spacecraft, the 46-bit reference frequency and the compression time; ramp start and end times, rates and start frequencies. On top of it, a ramped 2-way X-band Doppler model (integrated uplink ramps, round-trip light time with Shapiro and troposphere, DSN stations through the IAU 2006/2000A CIO chain and IERS EOP, station time) and a Mars-orbiter batch orbit determination (GMM-3 field, third bodies, SRP, drag).
Oracle
The PDS3 label that the DSN navigation data-conditioning process produced with the file (NASA PDS Geosciences Node, MRO-M-RSS-1-MAGR-V1.0), checked against bars pre-registered in xval/mro-odf/PREREGISTRATION.md before the reader existed. MODELLED because one registered bar (the end-of-file table's rows, 121 in the file against 120 in the label) missed; the label is internally inconsistent there. Stage 2 (xval/mro-od/PREREGISTRATION.md, amendments A1 and A2 before the scored result): the Doppler predicted from the MRO navigation team's reconstruction (mro_psp5, NAIF) matches the observed ramped 2-way Doppler to 12.5 mHz on passes not used in development, bar met; the registered orbit determination failed on the scored arc (normal matrix not positive definite) and produced no orbit, so the orbit bars were not met. Run 2 (xval/mro-od/PREREGISTRATION-RUN2.md, a damped estimator, the same arc and bars, a scoring amendment R2-A1 disclosed in docs/revisions) converged: the estimated orbit is within 0.36 km RMS (0.53 km maximum) of the reconstruction, bars met, but the post-fit Doppler RMS is 48.6 mHz against the 30 mHz bar, not met, so the row stays MODELLED. WHAT IS NOT: no estimated orbit has met every registered bar; ionosphere, solar plasma and most of the antenna offset (a 1.5 m radial offset is fitted on the development pass) are not modelled.
Tests
tests/mro_odf_reader_label_oracle.rs (a real MRO ODF, 2007-304/305, against its PDS label: record count, group positions, stations, start and stop times and data types hold; the end-of-file table's row count misses by…
ValidatedClosed-loop MRP feedback attitude control with reaction wheelsaocs (AocsConfig,required_torque,motor_torques,rates,rk4_step,simulate,mrp_from_quaternion); aocs_kind (AocsScenario)
Capability
Inertial-pointing attitude control of a rigid hub with three or more balanced reaction wheels: the MRP feedback law L_r = -K sigma_BR - P omega_BR (Schaub and Junkins), minimum-norm wheel motor torques G_s u = -L_r clipped to each wheel's limit, the coupled hub and wheel dynamics with the hub inertia carrying the wheels' spin inertia, quaternion kinematics, RK4 with a zero-order hold. Runnable as the aocs scenario kind (true-state loop).
Oracle
Basilisk 2.9.1 (AVSLab, ISC; the bsk wheel run as a separate tool, never linked or vendored): spacecraft hub, reactionWheelStateEffector (BalancedWheels, torque limit on), simpleNav, inertial3D, attTrackingError, mrpFeedback (no integral term) and rwMotorTorque at dt = 0.1 s with its RK4, against Kshana's own implementation on three cases (three orthogonal wheels; a four-wheel pyramid; a saturating four-wheel case), bars pre-registered in xval/aocs-basilisk/PREREGISTRATION.md before the code and the run. All bars held on the first run. WHAT IS NOT: sensor noise in the loop, integral feedback, tracking guidance, wheel friction, jitter, speed limits and power, any real spacecraft's gains; the comparison checks the implementation of the law and the dynamics against an independent one, not a flight controller.
Tests
tests/aocs_basilisk_oracle.rs (3 cases, 601 samples each: MRP worst 2.4e-14, bar 1e-9; body rate 1.6e-15 rad/s, bar 1e-9; wheel speed 2.9e-11 rad/s, bar 1e-6; applied motor torque 1.2e-13 N m, bar 1e-9; t = 0 torques…
ModelledMultiplicative EKF attitude estimation from a gyro and a star trackeraocs_mekf (Mekf,GyroSim,attitude_error,spd_solve6); aocs_kind (estimator)
Capability
The Lefferts-Markley-Shuster multiplicative EKF: body-frame attitude error and gyro bias states, propagation with the gyro rate less the bias estimate and Farrenkopf's discrete process noise (angle and rate random walk), star-tracker updates with the residual 2 vec(q^-1 q_st) in Joseph form, and the multiplicative reset; a Farrenkopf discrete gyro and star-tracker simulator. The aocs kind can close the control loop on its estimate.
Oracle
Internal consistency: the filter's normalised estimation error squared against truth simulated from the same sensor models over Monte Carlo seeds, and convergence of the attitude and bias estimates. WHAT IS NOT: no external estimator or flight data has been compared (Basilisk 2.9.1 has no MEKF module); the sensor models are the textbook ones, without misalignment, scale factor or star-tracker field-of-view effects.
Tests
aocs_mekf::tests::mekf_is_consistent_over_monte_carlo_seeds (40 seeds, mean six-degree-of-freedom NEES over the second half within 4.5 to 7.5); aocs_mekf::tests::star_tracker_updates_bound_the_attitude_error…
Optimal three-axis attitude from weighted vector observations (star/sun/magnetometer directions): the deterministic two-vector TRIAD, Davenport's q-method (the exact Wahba-loss minimiser — the optimal quaternion is the largest-eigenvalue eigenvector of the 4×4 Davenport matrix K, solved by a symmetric Jacobi eigensolve), and QUEST (Newton/secant root of K's characteristic equation seeded at Σ weights, then Gibbs-vector quaternion recovery)
Oracle
scipy.spatial.transform.Rotation.align_vectors (SciPy 1.13; Virtanen et al., Nature Methods 2020) — the Kabsch/Markley SVD solution of Wahba's problem, a genuinely independent algorithm (SVD of the attitude profile matrix) and codebase from kshana's Davenport-K eigensolve / QUEST characteristic-root method, computing the SAME uniquely-defined Wahba-optimal attitude. On noiseless weighted vector-observation sets TRIAD, the q-method and QUEST reproduce scipy's optimal rotation to <1e-9 rad (compared via the sign-invariant attitude-error angle). The noisy Monte-Carlo 'optimal beats TRIAD' statistical claim stays MODELLED (no external oracle); point-direction unit-vector observations only — no sensor field-of-view/bias/temporal-correlation modelling, and QUEST is singular at 180° (the q-method covers that case)
Tests
wahba::tests (A(identity quaternion)=I; the Jacobi eigensolver satisfies Kv=λv and preserves the trace on a known symmetric matrix; K is symmetric); tests/wahba_reference.rs (TRIAD, Davenport's q-method and QUEST…
Square-law (non-coherent) acquisition detector over a code-phase × Doppler search: false-alarm probability (central χ² with 2M dof), the threshold achieving a target P_fa (χ² CDF inverted by bisection), and detection probability (non-central χ² with non-centrality 2M·ρ for per-cell post-correlation SNR ρ), with the generalized Marcum Q-function Q_M(a,b)=1−F_{χ'²(2M,a²)}(b²)
Oracle
scipy.stats.ncx2 / scipy.stats.chi2 (SciPy 1.13; Cephes/Boost) — an independent algorithm (continued-fraction non-central/central χ²) from kshana's incomplete-gamma series (raim::chi2_cdf / noncentral_chi2_cdf), computing the same uniquely-defined detection-statistics KERNEL: the generalized Marcum Q_M(a,b)=ncx2.sf(b²,2M,a²), the detection probability P_d, the false-alarm probability P_fa and the P_fa→threshold inversion, matched to ~1e-6 — the same independence basis as the Validated 'RAIM/ARAIM integrity statistical kernel' row. KERNEL only: the per-cell CFAR cell-averaging and code/Doppler-bin straddling loss stay MODELLED
Tests
acquisition::tests (Marcum-Q central case Q_1(0,b)=exp(−b²/2); the P_d=Q_M(√(2Mρ),√γ) identity; P_fa↔threshold round-trip across M and P_fa; ROC monotonicity — P_d rises with SNR, falls with threshold, P_d→P_fa as…
Integer least-squares ambiguity fixing the LAMBDA way: a volume-preserving integer (Z) decorrelating transform (integer-Gauss size reduction of the L D Lᵀ factor) + an exact Schnorr–Euchner depth-first branch-and-bound integer least-squares search, with the ratio test on the two best candidates (the validated claim); plus the closed-form bootstrapped success rate P_s=∏(2Φ(1/(2σ_{i|I}))−1), checked only internally and outside the validated claim
Oracle
RTKLIB v2.4.2-p13 lambda() (its lambda.c source file, T. Takasu, BSD-2-Clause), an independent LAMBDA reduction and MLAMBDA search compiled from C and run as a tool, on 300 random covariances (n 2 to 10): integer least squares has a unique minimiser, so the fixed integers must be identical (0 mismatches) and the two best squared norms, which feed the ratio test, agree to 1e-9 relative (measured 5.6e-13); tolerance fixed before the comparison. Validated claim: the ILS solution and its two best norms. RTKLIB does not compute the bootstrapped success rate, which stays checked internally against a Monte-Carlo of the rounding process; the Z-transform invariants and the brute-force enumeration remain as supporting checks. Integer-Gauss decorrelation only (the reordering permutations of the full LAMBDA reduction speed the search but change neither the exact ILS answer nor the bootstrapped rate)
Tests
lambda::tests (L D Lᵀ reconstructs Q); tests/lambda_reference.rs (the Z-transform is unimodular |det Z|=1 with Q_z=ZᵀQZ SPD, det-preserving, and lower total off-diagonal correlation; the Schnorr–Euchner ILS matches…
Hyperbolic-flyby geometry (a=−μ/v∞², e=1+r_p·v∞²/μ, turn angle δ=2·asin(1/e), impact parameter |B|=|a|·√(e²−1)), the B-plane Ŝ/T̂/R̂ aim-point frame and B·T̂/B·R̂ decomposition, and a patched-conic gravity assist (v∞-magnitude conserved, direction deflected by δ, heliocentric Δv=2·v∞·sin(δ/2) at no propellant cost) with the Tisserand parameter T_P=a_P/a+2√((a/a_P)(1−e²))cos i
Oracle
GMAT R2026a (Apache-2.0, run as a separate program) on 192 hyperbolic Earth states: B.T, B.R and |B| within 1e-6 km (worst 1.35e-8 km), the turn angle and both asymptotes within 1e-9 rad (worst 9.3e-15 rad); the assist velocity change from GMAT C3 and asymptotes within 1e-9 v_inf (worst 5.6e-14); heliocentric SMA, ECC and INC on 350 Sun-centred states before and after the assist within 1e-9 relative and 1e-12 rad (worst 2.2e-14 and 4.1e-14 rad); the Tisserand parameter against sbpy 0.6.0 (BSD-3) on GMAT elements within 1e-9 (worst 1.3e-15) and against 3 - C3/v_c^2 (8.9e-16); and the Tisserand closed form against Kasuga and Jewitt (2019) Table 8.1 for 12 objects within the printed-rounding bound (worst 0.39 of it; a P1 published worked value). Pre-registered (0e2bb455; heliocentric repair 329f2252); amendment 3e974bb4, written after the oracle files were generated and before any Kshana value, widened the inclination input filter from 1e-2 to 1e-3 rad, keeping more and worse-conditioned states. MODELLED scope: patched-conic two-body flyby on a circular planetary orbit, no sphere-of-influence transition, encounter perturbations or ephemeris
Tests
bplane::tests (the flyby scalars satisfy the closed forms with two agreeing |B| identities and |B|>r_p; the turn angle decreases with periapsis radius and hits the δ→0 / δ→π limits; deflection preserves v∞ speed and…
ValidatedCRPA anti-jam array beamformingcrpa
Capability
Controlled-reception-pattern antenna nulling: complex array steering vectors a(û)=exp(j·k·pₙ·û), deterministic null-steering (unit gain toward the SV, exact nulls toward up to N−1 jammers via the minimum-norm w=Aᴴ(AAᴴ)⁻¹b), and MVDR adaptive weights w=R⁻¹a_sv/(a_svᴴR⁻¹a_sv) that self-steer deep nulls onto strong interferers (with a from-scratch complex linear-algebra kernel)
Oracle
numpy.linalg / scipy.linalg (LAPACK zgesv complex solve) — an independent linear-algebra codebase from kshana's from-scratch complex kernel, computing the SAME uniquely-defined weights: MVDR w=R⁻¹a_sv/(a_svᴴR⁻¹a_sv) and minimum-norm null-steering w=Aᴴ(AAᴴ)⁻¹b, plus the resulting array-response gains, matched to ~1e-9 — the same class of oracle as the Validated DOP (gnss_lib_py) and Fisher-information (numpy) rows. MODELLED narrowband far-field identical-isotropic-element array — no mutual coupling, per-element mismatch, finite-bandwidth (STAP), or steering-vector estimation error
Tests
crpa::tests (complex arithmetic identities incl. z/z=1, |exp(jθ)|=1, zᴴz=|z|²; deterministic null-steering gives exact unit SV gain and <1e-9 jammer nulls on a 4-element ULA with 3 jammers; an N-element array rejects…
The five-coefficient IEEE-1139 PSD↔Allan-variance conversion S_y(f)=Σh_α f^α → σ_y²(τ) (white/flicker PM, white/flicker/random-walk FM), the flicker-FM floor σ_y=√(2 ln2·h_{-1}) as a first-class fittable term, and a non-negative FM-family fit in the {(2π²/3)τ, 2 ln2, 1/(2τ)} basis that recovers the floor the drift basis {1/τ,τ,τ³} cannot represent
Oracle
allantools 2024.06 (A. Wallin; Kasdin & Walter 1992 / Vernotte 2015 coefficients) — an independent third-party codebase and citation lineage computing the SAME uniquely-defined IEEE-1139 PSD→Allan conversion σ_y(τ) for each of the five power-law noise types. kshana's powerlaw::allan_deviation matches the allantools closed forms (white/flicker-FM, RW-FM and white-PM to <5e-9 relative; flicker-PM to ~1e-5 — the genuine independence signal, kshana using the NIST-SP-1065 tabulated 1.038 constant vs allantools' full 3γ−ln2), and the flicker-FM floor √(2 ln2·h_{-1}) matches allantools across the ladder — the same bar the Validated MDEV/TDEV/Theo1/MTIE rows clear. MODELLED stationary power-law model — no deterministic-drift (τ²) term; per-device coefficients / measured floors stay Modelled
Tests
powerlaw::tests (each pure noise type shows its signature ADEV log-log slope — white FM −½, flicker FM 0, random-walk FM +½, white PM −1; the flicker-FM term is a flat floor equal to √(2 ln2·h_{-1}) across five decades…
Serialise and parse the CCSDS 502.0 OEM COVARIANCE_START…COVARIANCE_STOP block — the 6×6 position/velocity covariance written as its 21-element lower triangle (canonical CX_X … CZ_DOT_Z_DOT order) with the optional COV_REF_FRAME, complementing the existing OEM state-vector writer/parser so orbit-determination and filter covariances round-trip in the standard format
Oracle
oem 0.4.5 (MIT, https://pypi.org/project/oem/) — the SAME independent CCSDS-502 parser trusted by the Validated 'CCSDS OEM interoperability' row, a completely separate codebase (its own KVN tokenizer, covariance-section state machine and numpy matrix assembly). kshana emits a full OEM segment carrying its unlabelled bare-number lower-triangular COVARIANCE block; oem parses it and reconstructs the symmetric 6×6, matching kshana's input covariance element-for-element to f64 round-off — a genuine library-vs-library interchange round-trip (compared against oem's reconstruction, not a kshana re-parse), closing the row's previous 'no third-party fixture' gap
Tests
oem::tests (a symmetric PD covariance round-trips KVN→parse to f64 round-off and stays symmetric; the block emits exactly the 21 lower-triangular entries — i+1 per matrix row — with/without COV_REF_FRAME; a block with…
Position-error-vs-coast-duration budget built from IMU coefficients, each error source (accelerometer bias, gyro bias, velocity random walk, angle random walk, scale factor cruising and under sustained specific force) propagated through a nine-state local-level INS error model with the Schuler feedback (inertial::coast::ErrorDynamics after Groves 2nd ed. ch. 14: attitude, velocity and position errors, the vertical channel with its gravity gradient, WGS-84 radii and normal gravity, a level platform heading north at the configured speed, latitude_deg and height_m; systematic terms by an augmented matrix exponential, random walks by Van Loan), the flat-Earth monomials kept only as leading_order_m; contributions combined under a stated rule (rss / linear-sum / deterministic-sum-with-stochastic-rss); the coast durations reaching caller-supplied position thresholds (10 m and 50 m by default) located by a forward scan for the first crossing (first_crossing_s, since the Schuler-bounded curves are not monotone), with the dominant source at each crossing; and a TRN-bounded mode giving the largest terrain-fix interval that holds each threshold. Runnable as the ins-trn-coast scenario kind
Oracle
NaveGo v1.4 (commit 24d9488, LGPL-3.0), run as a separate program under GNU Octave 8.4.0: free-inertial coast runs with the same IMU error profile, with one corrected numerical dead band in its quaternion update (qua_update.m: the guard on a rate below 1e-8 rad/s replaced by wnorm == 0, applied to a run-time copy and asserted to match once; with the guard in place a slowly drifting platform attitude froze and cut the Schuler feedback). Pre-registered (f18b591f) with the corrected generator before it was run; the engine (cce1c47a) was committed before any corrected output was opened; same 5 % bar, 48 points, terms, seeds and statistic as round 1. Result: 48/48 within 5 % (worst +3.70 % velocity random walk at 1200 s, +3.33 % angle random walk at 1800 s); at 3600 s NaveGo/Kshana give 2375.4/2376.2 m for the accelerometer bias and 130 885/131 675 m for the gyro bias. Zeroing the Schuler transport-rate feedback fails the test (+695 % at 3600 s). Disclosed: single-channel Schuler predictions and the round-1 NaveGo values were known; the model has no fitted parameter. Scope: the per-term error curves at Tactical grade, 45 deg N, level and north-pointing, up to 3600 s; the crossings, the combination rule and the TRN mode are deterministic functions of those curves. The IMU class coefficients (Groves 2013 Table 4.1 band figures) and the TRN fix residual stay MODELLED inputs. Round 1 (superseded): the flat-Earth monomials, now leading_order_m, overstate beyond 600 s (2.2x to 8x at one hour). 0.30 revision: every ins-trn-coast output moves (the shipped scenario TRN peak 52.932679 -> 44.711834 m)
Tests
tests/ins_coast_schuler_navego_oracle.rs::coast_error_model_matches_the_corrected_navego_runs (48 points, six error terms at 30 s to 3600 s, Tactical grade, 45 deg N, level and north-pointing: every point within 5 %…
One run takes a contact plan (a list of aos_s/los_s windows, each asking the aperture for navigation or communications — the same window vocabulary passes::predict_passes emits, converted by aperture_duty::windows_from_passes), an aperture count and an explicit arbitration policy, and returns the navigation duty, the communications duty, the idle duty and the per-session outage. The policy is an input with a documented default (navigation-priority; also communications-priority and non-preemptive first-come-first-served), and the report carries the resolved policy, its full definition and the duty and outage definitions, because a duty figure quoted without its arbitration rule is not reproducible. The schedule is an exact interval sweep over the window boundaries, so a window that ends exactly where the next begins never contends, and navigation, communications and idle aperture-seconds are accumulated independently in that one sweep. Runnable as the aperture-duty-cycle scenario kind
Oracle
Internal consistency, from three quantities computed by different expressions in the same sweep and then required to close. Navigation aperture-seconds, communications aperture-seconds and idle aperture-seconds are accumulated separately — idle from (apertures − |served|) per elementary interval, never as a remainder — and their sum is asserted against apertures × horizon_s, which a mis-bracketed boundary, a double-counted interval or a dropped one breaks immediately; the per-session served times are separately required to add back to the two service totals. The scheduler's structural properties are measured rather than assumed: served time is asserted non-decreasing in the aperture count over 200 pseudo-random plans for all three policies (not obvious for the non-preemptive policy, where an extra aperture changes who holds what at every later boundary), and the boundary arithmetic is pinned at the touching/one-second-overlap pair where an off-by-one would otherwise hide. The pass-predictor bridge is checked against passes::predict_passes output: with one aperture and no competing service the navigation aperture-seconds equal the predictor's own summed pass durations. No external oracle is claimed and none exists: the capability supersedes hand arithmetic rather than being checked against it, the bundled contact plan is illustrative rather than flown, and no operational scheduler publishes a plan-plus-answer pair with its arbitration rule stated — a duty computed under an unstated policy is not comparable to one computed under a stated one, and the policy dependence is measured (navigation duties of 100/150 vs 50/150 on the identical plan). Slew and changeover time, data volume, buffer state, energy and link closure are excluded in the report label rather than silently modelled
Tests
aperture_duty::tests (a window that ends exactly when the next begins does not contend, and moving it one second earlier costs exactly one second — the off-by-one guard; overlapping windows beyond the aperture count…
Lunar-native jammer-to-signal ratio, effective C/N₀ and loss of lock for a selenographic surface user under a lunar surface (or raised) jammer, over the illustrative public-source Moonlight/LCNS-class constellation. Composes the open jamming chain (j_over_s_db, effective_cn0_dbhz, rx_antenna_gain_db, lock_status, q_factor, nominal_cn0_dbhz, free_space_path_loss_db) with the open lunar sky geometry (lunar_service::LunarConstellation + topocentric, lunar::selenographic_to_mcmf); no geometry or radiometry is re-derived. Unlike the Earth jamming kind the signal leg is not a fixed received power: each satellite's isotropic received power is its own link-budget EIRP − FSPL(slant range), so J/S varies satellite by satellite with lunar range and elevation. Emits ONE ROW PER VISIBLE (epoch, satellite) LINK as JSON and as a *.table.csv artifact, with both received powers the J/S is the difference of printed alongside it; aggregate figures of merit sit beside that table, never in place of it. Runnable as the lunar-jamming scenario kind
Oracle
Composition cross-check against an independent in-repo code path: every per-link J/S is re-derived as the difference of two link budgets computed with linkbudget::received_signal_power_dbw, whose free-space loss is the single expression 20·log10(4πRf/c) rather than the three-term sum jamming::free_space_path_loss_db uses — the two agree to < 1e-9 dB on every row (measured worst case 7.11e-14 dB, i.e. float round-off). The underlying interference chain it composes is separately externally anchored in tests/gnss_denied_jamming_resilience_reference.rs (independent numpy/scipy re-derivation of J/S and effective C/N₀ pinned to Kaplan & Hegarty §9.4, plus JammerTest 2024 measured C/N₀, Zenodo 10.5281/zenodo.15910563); the lunar geometry it composes is the lunar_service/lunar geometry with its own oracles. Closed-form identities checked here: a halved jammer standoff adds exactly 20·log10(2) dB to every row; J/S is exactly invariant to the user-antenna boresight gain while C/N₀ moves by exactly that gain; an elevation mask changes no surviving row bit-for-bit. The row itself stays ReferenceImpl/Modelled: no measured lunar jamming campaign exists to validate against, and the constellation is an illustrative public-source Moonlight/LCNS-class geometry, not a flown ephemeris
ModelledLunar denial contour with an uncertainty band from the measured C/N₀ spreadlunar_jamming (denial_js_db,range_for_free_space_path_loss_m,Cn0Distribution,DenialContourPoint,ContourBand,DenialContour; inverting jamming::effective_cn0_dbhz,jamming::j_over_s_db and jamming::free_space_path_loss_db)
Capability
The lunar-jamming report no longer rests its denial contour on one scalar. It emits the full measured wanted-signal C/N₀ distribution over the per-satellite table — n, min, p05, p25, median, p75, p95, max, mean, sample stdev, and the sample's own asymmetry (p95 − median) − (median − p05) — beside the per-link rows, which are unchanged. The contour is then reported at each of those order statistics under BOTH denial criteria the engine recognises, never one in place of the other: the incumbent power-ratio criterion (J/S = 30 dB, the same threshold attack_surface and tracking_loop use, and the criterion behind the released lunar link-jamming table's denial column), and the loss-of-lock criterion (the effective C/N₀ falling to tracking_threshold_dbhz), which is the criterion this report's own links[].status column is scored with. Each contour point is given on both axes of the denial plane — the jammer EIRP required at the scenario's standoff, and the denial standoff at the scenario's EIRP — so the band is an interval on the same axes a contour plot is drawn on. The band edges ARE contour(p05) and contour(p95): the same closed-form map applied to the sample's own quantiles, not a sigma fitted to the sample and not median ± k·stdev, and the report says so in a band_definition string beside the numbers. stdev is emitted for continuity and is used by nothing. A quantile whose C/N₀ is already at or below the tracking threshold has no finite denying J/S; those columns are emitted as null with a counted reason rather than as an infinity or a clamped radius
Oracle
Internal consistency against the engine's own forward functions, which is all this capability can honestly claim. Every contour point is a closed-form inversion, and each inversion is checked by pushing its answer back through the FORWARD function the report's own rows were scored with: the J/S column through jamming::effective_cn0_dbhz (which must return tracking_threshold_dbhz to < 1e-9 dB-Hz), the standoff column through jamming::free_space_path_loss_db and jamming::j_over_s_db (the same J/S to < 1e-9 dB), the EIRP column through jamming::j_over_s_db at the scenario's own standoff. The contour is further checked to be the BOUNDARY of the denial set rather than a point inside it — lock_status loses lock on a strict inequality, so the verdict is measured to flip from DEGRADED to LOST across 1e-4 dB of J/S either side of the contour, which an off-by-an-epsilon contour would fail. The strongest check is per-link rather than per-quantile: applying the same map to every row's own C/N₀ reproduces the report's status column exactly — 0 disagreements on 38 of 38 rows at the documented operating point (kind = lunar-jamming, every input default except jammer.range_m = 25 000 m), where 26 rows are LOST and 12 are not. Monotonicity is measured, not assumed: 20 001 samples across [tracking_threshold + 1e-3, 80] dB-Hz, 0 violations of a strictly decreasing standoff and a strictly increasing required power, so the quantile ordering of the band is demonstrated rather than asserted. ExternalDataset is declined and Validated with it: no measured lunar jamming campaign exists to compare a denial radius against, the constellation is an illustrative public-source Moonlight/LCNS-class geometry rather than a flown ephemeris, and the contour's inputs (EIRP, jammer power, antenna gains, noise temperature) are representative magnitudes. ReferenceImpl was considered and declined too — the inverse and the forward expression are the SAME algebra read in two directions, so the round trip catches transcription and sign errors but is not an independent implementation; calling it a cross-check would be the borrowed-independence move the matrix invariants exist to prevent. What the band buys is measured and quoted rather than asserted: at that operating point the single-scalar contour is 27.402916 km and puts 25 km on the denied side for all 38 rows, while the band 21.256108 .. 29.921360 km contains the operating point and therefore reports the split the table actually shows. The two criteria are also measured to disagree — the power-ratio band (38.186527 .. 53.691675 km) does not contain 25 km at all — which is why both are printed
The hybrid-optical-rf report states what the cross-modality chi-square monitor can actually DETECT, not only that it passes a fault-free case: the minimum detectable bias per monitored axis (east, north, up, clock) at the scenario's stated false-alarm and missed-detection probabilities, the detection-power curve either side of it, the noise-free bias multiple at which the realised statistic first crosses the threshold, and the time-to-detect for a bias ramp at a stated rate. Faults are also injected through the real monitor and the statistic it returns is reported alongside the analytic prediction
Oracle
Two internal oracles, no external dataset. (1) Closed-form round trip: the minimum detectable bias is produced by inverting the non-central chi-square tail on the non-centrality (raim::pbias) and is verified by feeding the resulting non-centrality back through raim::noncentral_chi2_cdf at the monitor's own threshold, which must return P_md. (2) Injection vs analysis: a bias is written into the RF estimate of one axis and cross_raim::run_cross_raim is re-run, so the statistic compared against the analytic non-centrality is the one the monitor computed, not a re-derivation. A seeded Monte-Carlo of the statistic itself is carried in the tests as a third check; it is deliberately NOT in the report, because a sampled estimate would be slower and not reproducible bit-for-bit. ExternalDataset is declined: the quantity is the detection power of THIS monitor at THIS scenario's sigma allocation, and the sigma magnitudes the MDB is scaled by are Modelled representative inputs, not measurements. ReferenceImpl was also considered and declined — the Monte-Carlo samples the same statistic the analysis describes, so it catches transcription and coefficient errors but is not an independent implementation of the monitor. 0.30 round 2 (pre-registered 0913a98a), blocked (stays MODELLED): SciPy 1.18.1 chi2/ncx2/brentq reproduce the report threshold (within 5.4e-12), per-axis MDB, power curve, crossing multiple and ramp time-to-detect (within 5.3e-10) on five configurations; the injected-fault statistics have no independent oracle and the sigma magnitudes are Modelled inputs, so the scoping awaits an owner decision
Tests
hybrid_integrity::tests (the detection-power curve runs from the false-alarm rate at zero fault to 1 − P_md at the minimum detectable bias; the minimum detectable bias fed back through noncentral_chi2_cdf at the…
ModelledPost-handover covariance re-growth against the alert limithybrid_integrity
Capability
The hybrid-optical-rf report exposes the filter's process-noise model and states how long the post-handover solution stays inside the alert limit: for both handoff directions it carries the per-axis covariance the handover left behind, propagates it forward under a stated random-walk process noise, and reports the time at which the coverage-scaled 1σ reaches the horizontal, vertical and timing alert limits, which limit binds first, the variance doubling time constant, and a sampled coast profile that brackets the crossing. Process-noise PSDs, the coverage factor and the alert limits are all inputs with documented defaults
Oracle
No external oracle. The crossing is a closed-form solution of the random-walk variance growth, and the tests recompute it independently from the values the report itself publishes rather than from the emitter's internals. The starting covariance is pinned bit-for-bit to the trace the existing handoff block already reported, so the coast cannot propagate a covariance nobody else saw. The process-noise PSDs are MODELLED representative inputs and the crossing times scale directly with them, which the scaling test states as a measured property rather than a claim. ExternalDataset is declined deliberately: the coast time is a direct function of two Modelled PSDs and of alert limits that are themselves representative, and there is no measured lunar optical/RF handover coast to compare against. An external oracle here would need a published post-handover covariance-growth or holdover-accuracy curve for a comparable receiver with its process-noise model stated, checked to a tolerance
Tests
hybrid_integrity::tests (the coast starts from exactly the covariance trace the handoff block reports, for both directions, so the replayed diagonal cannot drift from the reported handover; the reported crossing time…
ModelledOff-boresight antenna pattern in the lunar geometry exportlunar_service,antenna
Capability
The per-satellite geometry export carries, per (epoch, satellite), the off-boresight angle AT THE SATELLITE and the transmit gain toward the site from the real uniformly-illuminated circular-aperture (Airy) pattern, and reports the in-beam count under that pattern BESIDE the in-beam count under the symmetric gain-to-beamwidth approximation (θ_3dB[deg] = √(31000/G_lin)), with the difference emitted as an explicit correction in links, in satellites per epoch, and as the worst single epoch. Both implied aperture efficiencies of the approximation are emitted (0.641 against the 70·λ/D degrees rule it is quoted with, 0.920 against a uniform circular aperture) so the efficiency it silently assumes is stated rather than inferred. Off by default: the block appears only when an export site and an aperture are both given
Oracle
Mixed, and separated rather than pooled. The PATTERN underneath is externally validated and keeps its own row (a scipy.special.j1 fixture to < 0.05 dB in tests/validate_p1_orbital_footprint.rs, plus here the published Airy half-power abscissa x = 1.61634 located by bisection rather than assumed). The GEOMETRY is checked against closed-form triangle trigonometry written as a different expression from the dot product under test. The CONTAINMENT of the real beam by the approximate cone is derived algebra (28.019/√η vs 29.479 degrees per λ/D), so the correction can only be non-positive for η ≤ 0.9035. But the in-beam COUNTS themselves have no external oracle: no published table gives how many satellites of a Moonlight/LCNS-class shell hold a south-polar site inside a given dish's half-power beam, and the constellation is an illustrative public-source approximation, not a flown ephemeris. The measured disagreement at the documented working point — 0 links in beam under the real pattern against 28 under the approximation, −2.33 satellites per epoch, worst epoch 3, on 76 evaluated links — is therefore a MODELLED finding about the approximation, pinned as a regression literal (the nearest row sits 0.069° from either beam edge, four orders of magnitude above any last-digit disagreement), not an externally validated coverage number. Labelling the row ExternalDataset on the strength of the pattern's own external anchor would be borrowed validation, which is the move the matrix invariants exist to prevent. 0.30 round 2 (pre-registered 67d5f30c), blocked (stays MODELLED): SPICE geometry and SciPy Bessel functions reproduce the off-boresight angle (worst 3.0e-5 deg), Airy gain and in-beam flags on 6 336 rows of flown-orbiter states and the documented working point; but on flown geometry the flags split by satellite identity (the nearest link is 1.66 deg from the beam edge), the 5 deg visibility is checked as a flag, not a numeric elevation, the symmetric-approximation flag applies the 31000/G rule of thumb to a validated angle without validating the rule, and the implied efficiencies have no oracle; the scoping awaits an owner decision
Tests
antenna::tests and lunar_service::tests (the half-power crossing located by bisection on the pattern matches the published Airy x = 1.61634 and yields the exact 1.02899·λ/D width, recording that the conventional 1.02…
ModelledTracking-loop loss of lock and spoof pull-in under interferencetracking_loop (composing sdr,jamming)
Capability
Loss of lock computed from loop dynamics instead of a power ratio: carrier (Costas) and code (non-coherent early/late) 1σ thermal jitter against C/N₀ with the squaring loss, against the stated rules 3σ_PLL + θ_e ≤ 45° and 3σ_DLL + ramp lag ≤ d/2 chips; drop and re-lock C/N₀ thresholds with the binding loop named and the hysteresis DERIVED from the wider pull-in bandwidth rather than asserted; the declared time to lose lock from a two-threshold lock detector with confirmation dwells, reported separately from the physical phase-escape time (Viterbi mean time between cycle slips, in log₁₀ s because it spans hundreds of decades); the largest code slew and carrier Doppler rate the victim's loops can follow; and the DENIAL RADIUS the loop dynamics imply reported alongside the existing power-ratio radius with their signed difference as its own named field. Runnable as the tracking-loop scenario kind
Oracle
The engine's own sdr correlator (sdr::correlate / synth_if / CaCode) stepped forward on seeded synthetic IF at a calibrated C/N₀ — a separate code path reaching the same numbers by numerical correlation of sampled IQ rather than by an algebraic jitter expression, so it is a different route and not a restatement — plus standard tabulated modified-Bessel I₀ values for the cycle-slip term. Loop theory per Kaplan & Hegarty ch. 8 and Viterbi/Gardner for the slip time. The row stays ReferenceImpl/MODELLED: the cross-check lives in this same codebase, and the bessel_i0 check validates one special function rather than the tracking model, so promoting the row on that basis would be self-serving labelling. ExternalDataset would need a recorded raw-IF dataset with ground-truth C/N₀ and an annotated loss-of-lock instant (TEXBAT/OAKBAT class); none ships in this tree, and even with the IQ those datasets publish no per-epoch loop-state truth, so a declared loss-of-lock time could only be validated against some other receiver's lock detector, which is a modelled choice and not an oracle. The loop bandwidths, integration time, correlator spacing, pull-in ratio, dwells, jammer power and antenna gains are representative band figures, not a datasheet. 0.30 round 2, GNSS-SDR 0.0.19 (GPL-3.0, run as a program) on independently generated IF (pre-registered 468e7f10, amendments to the measurement method disclosed), a finding (stays MODELLED): closed-loop PLL and DLL jitter agree within 8 % at 35 to 45 dB-Hz, the 10 Hz/s dynamic stress within 1 % and the first-order code ramp lag within 2 %; at 30 dB-Hz the PLL jitter is 21 % higher and the DLL jitter 32 % lower, the measured 15-degree-rule thresholds sit 1.5 dB above the engine, and the second-order loop slips 0.5 to 2 decades more often than the first-order slip-time bound. 0.30 revision: the Costas slip time now uses the loop SNR rho/4 (half-cycle slips); the default jitter_table log10_mean_time_to_cycle_slip_s falls from 11.58 to 2.10 at the 25.47 dB-Hz threshold
Tests
tracking_loop::tests (open-loop Costas discriminator jitter against sdr::correlate stepped forward on seeded synthetic IF, 3 pooled noise realisations × 900 epochs, agreeing to 0.63% over 35-45 dB-Hz, with the…
ValidatedLunar-VLBI station-coordinate covariance from a tracking schedulelunar_vlbi_fim (composing lunar_vlbi,fim,cio,lunar_frame,frames)
Capability
Delay partials accumulated over a schedule of baselines × epochs into a Fisher information matrix, inverted to the station coordinate covariance and the per-coordinate station sigma, replacing an assumed isotropic equipartition link from a scalar delay precision. The state carries Earth-fixed (ITRS) station coordinates, so the Jacobian is the inertial partial rotated by each epoch's GCRS→ITRS matrix and Earth rotation is what makes those coordinates observable — the report MEASURES the Earth-fixed line-of-sight sweep and the beacon declination rather than assuming them. Rank, datum defect, condition number, the information spectrum and the free-network null space are emitted on every run, and under a rank deficiency the headline sigma is published as NULL with a status rather than read out of a near-singular inverse. The equipartition value c·σ_τ·√(g/N) for the SAME schedule is printed beside the computed one with their ratio, together with the isotropic trace bound √(p/trace(M)) that AM-HM makes a hard floor. Runnable as the lunar-vlbi-fim scenario kind
Oracle
An independent in-repo route. Every Jacobian row is re-derived by central finite difference of lunar_vlbi::vlbi_delay_s evaluated from the state's own Earth-fixed and Moon-body-fixed coordinates rotated forward through the frame chain — a path that computes no derivative and shares no expression with the analytic partials — and the information matrices the two routes build agree to < 1e-6 of the largest diagonal. The linear-algebra kernel this composes (information_matrix, crlb, sym_eig) is separately externally anchored against numpy.linalg.eigh / numpy.linalg.inv in tests/fim_observability_reference.rs, and the delay observable carries lunar_vlbi's own delta-DOR far-field oracle; the row does not borrow either status. Closed-form identities checked here and labelled as such: σ² and 1/N scaling, and √(p/trace(M)) as an AM-HM lower bound the computed sigma cannot beat, which coincides with the equipartition value only on an isotropic geometry — which is why the measured ratio is exactly the anisotropy the assumption discarded. 0.30 external comparison (Library: NAIF SPICE CSPICE N0067 via spiceypy 8.2.0, MIT; DE440, ITRF93 earth_latest_high_prec, MOON_PA), pre-registered (aad03856) together with an engine-side model-error sensitivity run: every Jacobian row is rebuilt by central differences of converged Newtonian light-time differences from SPICE geometry (the light-time equation is a short script over SPICE positions; SPICE carries the geometry), with the schedule rebuilt from SPICE elevations. The schedule, rank, defect, information spectrum, condition number, station sigmas, headline, trace bound, equipartition ratios, free-network null space, line-of-sight sweep and beacon declination agree within the pre-registered 1 % / 0.1 deg, which sits above the engine known analytic-Moon error (3.2e-3 on sigmas, measured before the run). NumPy 2.3.5 redoes the linear algebra on the committed engine Jacobian to 2e-13 (P2). Validated as a Cramér-Rao bound for the stated reduced parameter set on the stated illustrative inputs: the Moon-centre ephemeris, clocks, troposphere and Earth-orientation parameters are held fixed in both legs, so it is not a real-campaign accuracy
Tests
lunar_vlbi_fim::tests (24 lib tests: the analytic Jacobian against a central finite difference of lunar_vlbi::vlbi_delay_s taken through a route that rotates the state's own coordinates forward and never touches a…
ModelledLunar-surface-point coordinate covariance from a VLBI delay schedule, kept distinct from the Earth-station onelunar_vlbi_fim (composing lunar_vlbi,fim,cio,lunar_frame,frames)
Capability
A third datum choice, all-stations-fixed, that holds every Earth station and estimates the BEACON's Moon-body-fixed coordinates alone — the configuration a lunar surface-point uncertainty is actually quoted for, since Earth station coordinates are an input to the delay model rather than an unknown of it. It exists because the station-level covariance and the surface-point covariance are DIFFERENT QUANTITIES separated by the lever arm ρ/B, and nothing previously stopped one being quoted against the other. The emitted beacon_link block carries ρ, the longest baseline, the lever arm, the surface-point form of the equipartition link c·σ_τ·(ρ/B)·√(g/N), the computed per-coordinate beacon sigma and their ratio — the ratio null rather than misleading whenever the beacon is unestimated or the matrix rank-deficient. B is taken as the LONGEST baseline present, the most favourable one, so the ratio can only understate. Runnable as datum = "all-stations-fixed" with estimate_beacon = true, which is refused with a message naming the missing input when the beacon is not estimated
Oracle
Closed-form identities, with no external oracle claimed or available. The lever arm is checked against ρ/B and the two equipartition forms against each other to 1e-15 — algebraic identities, so they catch a wiring error and nothing else, and are labelled as such. The substantive result is MEASURED rather than asserted: a delay from a fixed baseline constrains the beacon's DIRECTION, so the line-of-sight component reaches the information matrix only through the near-field range term, and the test reads the resulting anisotropy off the emitted spectrum instead of deriving it. AM-HM makes √(p/trace(M)) a hard floor, so an isotropic-equipartition link can only ever UNDERSTATE, and the computed-over-equipartition ratio is exactly the anisotropy it discarded. MEASURED on the Goldstone-like/Canberra-like pair whose chord is 10726.748 km: over 112 sampled days, 36 admit no mutually visible epoch at all at a 10° mask, the most ever mutually visible is 17 of a 49-sample schedule, and on the 73 viable days the three-component ratio runs from 113.9× to four orders of magnitude more, median 1513×, while restricting to the two transverse directions the line of sight does not starve gives 3.94× to 25.8×, median 7.8×. WHAT THIS ROW DOES NOT CLAIM: any of those figures as a property of a real campaign. It establishes that the surface-point quantity is now COMPUTED rather than assumed, and that it cannot be silently interchanged with the station-level one. Stays ReferenceImpl/MODELLED: no published lunar-VLBI surface-point covariance exists to validate against, the stations and beacon site are illustrative rather than surveyed, and the ephemeris, clocks, troposphere and Earth-orientation parameters are held FIXED, so this is a Cramér-Rao bound for a reduced parameter set and optimistic in its own right. The same quantity on the kernel path (beacon on the DE440 Moon) is validated on its own row, "…, kernel path"; on this analytic path the smallest beacon sigma is 3.9 % below that row's SPICE oracle, which is why this row stays MODELLED
Tests
lunar_vlbi_fim::tests (7 lib tests: every datum spelling round-trips through parse/as_str and an unknown one is refused; holding every station without the beacon is refused with an error naming estimate_beacon; the…
ValidatedLunar-surface-point coordinate covariance from a VLBI delay schedule, kernel pathlunar_vlbi_fim (epoch_geometry_with,schedule,schedule_jacobian); ephem_provider (KernelEphemeris); naif_kernel
Capability
The all-stations-fixed beacon covariance of the lunar-vlbi-fim scenario (every Earth station held, the beacon's Moon-body-fixed coordinates estimated alone) with the beacon placed on the JPL DE440 Moon read by the engine's own kernel reader (planetary_kernel_path; lunar_vlbi_fim::epoch_geometry_with, ephem_provider::KernelEphemeris). Only the Moon centre differs from the analytic path; the IAU 2015 body rotation, stations, Earth rotation and linear algebra are the same. The validated outputs are the observation set, the 3 x 3 information spectrum, rank and condition number, the three body-fixed beacon sigmas and their RMS, the lever arm and the computed-over-equipartition ratio, on the default schedule. Runnable as the lunar-vlbi-fim scenario with datum = "all-stations-fixed", estimate_beacon = true and planetary_kernel_path set
Oracle
Binding Library leg: NAIF SPICE Toolkit N0067 (spiceypy 8.2.0, MIT) on de440s, earth_latest_high_prec (ITRF93 with UT1 and polar motion) and the DE440 lunar frames, the beacon fixed in MOON_ME; every Jacobian row is a 3 km central difference of the converged Newtonian light-time difference (the M069 oracle's own light-time code, imported), never a Kshana partial; visibility rebuilt from WGS-84 elevations; NumPy 2.3.5 / SciPy 1.18.1 (BSD-3-Clause, LAPACK) eigh and inv. P2 leg: NumPy on the engine's committed Jacobian. Pre-registered (a4781fd) before the fixture or oracle existed, with bars of 1 % on eigenvalues, condition, sigmas, RMS and ratio, 1e-4 on the lever arm, exact observation set and rank; P2 1e-9. Disclosed: the engine-only comparison of analytic and kernel Moon (3.9 % on the y sigma) was run before the pre-registration, to choose the kernel path, and is stated there. Result: both legs agree, worst 1.7e-3 (beacon y sigma 0.081211 m against 0.081351 m). Mutation: transposing the beacon partial's body rotation fails both legs. Both sides read DE440, so agreement validates the geometry and linear algebra on DE440, not DE440. Outside the claim: a real campaign's surface-point accuracy (clocks, troposphere and Earth orientation held fixed; illustrative stations and beacon site), and the analytic default path (row above, which stays MODELLED: on the analytic Moon the y sigma is 3.9 % below this oracle)
The three terms a differential correction picks up between the epoch it is computed at and the epoch it is used at, each previously left as unmodelled headroom: reference-station survey error (correlated across satellites, baseline-independent, ~1:1 position transfer and provably not DOP-amplified), correction ageing (the frozen orbit-error vector re-projected onto the line of sight the engine's own propagator puts the satellite on one latency later, plus c·σ_y(τ)·τ clock drift from the calibrated power law), and uniform link quantization over an exact ±(orbit_err + clock_err) full scale with step²/12 variance. Reported in both the range and position domains, beside — never in place of — the existing residual and protection level, with a latency curve beside the single figure
Oracle
Closed-form identities and self-consistency: the uniform-quantizer step²/12 variance, the exact (GᵀG)⁻¹GᵀRG(GᵀG)⁻¹ covariance propagation collapsing to σ·PDOP for equal σ, the survey term's 1:1 transfer as a consequence of û_user ≈ û_ref, and switching-off quadrature closure. The PDOP cross-check against orbit::dop shares the [−û, 1] normal matrix, so it is a consistency check and is labelled as one rather than an independent oracle. No ExternalDataset is available or claimed: no lunar surface station has a published surveyed accuracy, no lunar differential-correction link has a published latency or message format, and the quantization result is a closed form rather than a measurement. RTKLIB — already the external-code oracle for the single-difference and weighted-least-squares kernel — could be driven with a deliberately mis-surveyed base to confirm the 1:1 survey transfer, but that is independent code over the same first-order line-of-sight algebra. MAGNITUDES ARE MODELLED: the survey default is this crate's own lunar frame-realisation allocation rather than a measured station, the latency and bit count are ILLUSTRATIVE inputs, and growth of the broadcast ephemeris error VECTOR itself is not modelled at all — stated in the emitted ageing-law string, because it is the term most likely to dominate a real link
Tests
lunar_dpnt::tests (a purely-additive guard whose key-set delta is exactly {correction_link, units} with every pre-existing value bit-identical against literals captured before the change; survey error does NOT…
ModelledLunar reference-station coverage and accuracy maplunar_station (over lunar_dpnt,lunar_service)
Capability
Monte Carlo sweep of the lunar differential-PNT model over a south-polar user grid and a day of epochs (the lunar-station-coverage kind): for a differential reference station at a known south-polar site, per point and over the region, the percentile corrected and uncorrected position error (3-D, horizontal, vertical), availability with a reason for each unavailable epoch (user_lt4, station_missing_sat), coverage against a threshold stated before the run, DO-229E protection levels, the coverage radius per azimuth sector, and the correction-link budget at the station and at the worst point. Not a statement about the performance of any real system; not affiliated with or endorsed by any agency
Oracle
Arithmetic re-computed by an independent numpy script on exported cases (scripts/gen_lunar_station_ref.py, numpy 2.5.4; pre-registered in tests/fixtures/lunar_station/PREREGISTRATION.md, committed alone in 2d43d3d5 before the first run): on 20 cases with every error term on, the visible sets and status are identical, the raw range errors agree to 1.4e-14 m, the position-error vectors and magnitudes to 1.4e-9 m against a 1e-6 m bar, and each point's p95 equals numpy.percentile bit for bit; the station corrections MISSED their 1e-9 m bar (worst 3.73e-9 m, 15 of 20 cases), and with them the corrected range errors: checked in 60-digit arithmetic, both sides are within double-precision rounding of the exact value, the survey term being a difference of two 1e7 m ranges whose last-place unit is 1.86e-9 m; recorded, not tuned. The kind reduces exactly to the lunar-differential-pnt kind at one point, one epoch and one trial. The error model, the decorrelation model and every threshold are this project's own or ILLUSTRATIVE; nothing here is a statement about any real system
Tests
tests/lunar_station_coverage_reference.rs (one_point_one_epoch_one_trial_reduces_to_lunar_differential_pnt: to 1e-12 m with and without receiver noise; zero_baseline_floor_is_zero_in_every_trial: 0 to 1e-9 m…
ModelledLunar terrain horizon masks from a stated profile or a polar DEMlunar_terrain (over skyline,lunar_service::topocentric_basis); lunar_station; lunar_service (coverage_masked,masked_visible)
Capability
Per-site azimuth-dependent horizon masks for the lunar-station-coverage kind (the reference station and every user point) and the moonlight-service-volume grid and its per-satellite export site (each export row gains clears_terrain), in the 0.36 skyline table format (skyline::Skyline, linear in azimuth with wrap at 360 deg; a stated profile and a DEM horizon at one site combine as their pointwise maximum): a stated [terrain] station_horizon or user_horizon labelled with its source, and/or a horizon ray-marched over a south polar stereographic DEM (the LRO LOLA gridded polar layout, read from a descriptor that fixes its size and SHA-256, or a latitude/longitude realdata::lola_dem tile with [terrain] dem_tile_path, echoed with its SHA-256) along great circles of the lunar sphere with exact curvature, from a stated mast height, at a stated azimuth step and range. A satellite counts at a site when it clears the flat mask_deg and the horizon at its azimuth (lunar_service::topocentric); an epoch the horizon leaves with fewer than four satellites at a point, where the flat mask gives four, is counted as the unavailable reason terrain_masked. The vendored DEM is a crop of the public-domain LOLA LDEM_75S_240M product, 5 x 5 block-mean resampled to 1200 m (data/lunar-dem/NOTICE.md). Absent, every site keeps the flat mask and every document is byte-identical
Oracle
Closed-form surfaces (InternalConsistency): the bare-sphere dip, a crater in a plateau on the sphere with an observer on and off its centre, and an analytic cone against its own rasterisation. On the real crop, an independent numpy ray-cast (scripts/gen_lunar_terrain_ref.py, numpy 2.5.3; pre-registered in tests/fixtures/lunar_terrain/PREREGISTRATION.md, committed alone in 4fa71559 before the script existed) agrees at five sites: site heights to 1.0e-11 m and all 1800 horizon elevations to 5.5e-10 deg against 1e-6 bars, truncated-ray counts exact; this checks the arithmetic, not the terrain. No published polar horizon or illumination figure has been reproduced: the vendored 1200 m crop smooths the sub-kilometre relief that published illumination studies resolve, so the DEM horizons are a MODELLED line of sight over a resampled real DEM, not a validated horizon. The station site and constellation stay ILLUSTRATIVE; nothing here is a statement about any real system
Tests
lunar_terrain::tests (bare_sphere_horizon_is_the_dip: the ray march reproduces -acos(R / (R + a)) to 1e-3 deg at every azimuth; a_cone_on_a_lat_lon_tile_matches_the_analytic_cone: the same cone sampled onto a 0.005 deg…
ModelledLunar reference-station multipath and receiver-hardware termslunar_receiver; lunar_station
Capability
The optional [receiver] section of lunar-station-coverage (LRS-02): elevation-dependent multipath sigma(el) = a + b exp(-el / el0) at the user and at the station (the RTCA DO-229 airborne functional form, its a = 0.13 m, b = 0.53 m, el0 = 10 deg as defaults; on the Moon every value is ILLUSTRATIVE), white at the user and first-order Gauss-Markov over epochs at the station (rho = exp(-dt / tau)); early-minus-late DLL thermal code noise sigma = L_c sqrt(B_L d / (2 C/N0) (1 + 2 / ((2 - d) T C/N0))); per-channel inter-channel biases constant over a trial, at the user and at the station. Station terms enter its corrections, user terms the raw and corrected measurements, each on its own seeded stream. Absent, every document is unchanged
Oracle
Closed forms and statistics of the model's own draws (InternalConsistency): the DLL formula against a hand computation, the DO-229 form, the zero-baseline difference of two receivers, the Gauss-Markov lag-1 correlation. No lunar multipath or receiver measurement exists to compare with; the values are ILLUSTRATIVE and nothing here is a statement about any real receiver
Tests
lunar_receiver::tests (multipath_is_the_do229_form; dll_noise_matches_a_hand_computation: 0.825279 m at 45 dB-Hz, 0.5 Hz, d = 1, 20 ms; bad_sections_are_refused); tests/lunar_station_ext.rs…
ModelledLunar reference-station ephemeris-error growth over the correction latencylunar_station (orbit_growth_range_sigma_m)
Capability
The optional orbit_err_growth_m_per_s input of lunar-station-coverage (LRS-03): a correction applied latency_s after it was computed misses the growth g latency_s of the broadcast orbit error, g isotropic of 3-D RMS rate, drawn per satellite and trial on its own seeded stream (range 1-sigma rate tau / sqrt(3)), on top of the frozen-error line-of-sight re-projection. Absent, every document is unchanged. A K5 addendum to the published-claims comparison, pre-registered alone, runs K5 at 0.001, 0.01 and 0.1 m/s (ILLUSTRATIVE)
Oracle
The growth term's closed-form sigma and the first-order propagation check (InternalConsistency). The K5 addendum (tests/fixtures/lunar_station_published/ADDENDUM-LRS03.md, pre-registered alone in f9ce1284): K5 fails at every rate, corrected p95 3-D 369.8 m, 2315 m and 22797 m against a 1.5 m bar (292.96 m recorded without growth), as stated before the run. No public source gives a lunar growth rate: ILLUSTRATIVE
Tests
tests/lunar_station_ext.rs (growth_absent_or_zero_is_the_frozen_error_run; the_growth_term_has_the_closed_form_sigma: the growth-on minus growth-off link error has sample sigma rate tau / sqrt(3) within 6 % over 1600…
The optional [solution] section of lunar-station-coverage (LRS-06): a height pseudo-measurement row up . dx = e_h of stated 1-sigma weighted against range rows, a three-satellite minimum allowed only with it, and a weighted least-squares solve over the last window_epochs epochs with a shared static position and one clock per epoch (a filter's answer for a static position with no process noise over the same data; the epochs' errors are treated as independent). Absent, the snapshot solve and every document unchanged. An addendum to the published-claims comparison with the sources' solution style, pre-registered alone
Oracle
The snapshot reduction and the solver's closed-form cases (InternalConsistency). The addendum (tests/fixtures/lunar_station_published/ADDENDUM-LRS06.md, pre-registered alone in 2b4c34b5) with a 1 m height constraint, three satellites and a 6-epoch window (ILLUSTRATIVE settings; the sources state a filter, a height constraint and three satellites but not their settings): K1 to K5 fail and K6 passes, 1 of 6 as in the recorded run; the R300 corrected p95 3-D error falls from 280.6 m to 2.11 m (horizontal 0.90 m, not scored) and coverage stays 0. The window's independence assumption favours it. Not a statement about any real system
Tests
lunar_solution::tests (one_epoch_without_height_is_the_snapshot_solve: equals lunar_dpnt::position_error_vector_from_range_errors to 1e-8 m, a loose height row changes nothing, a tight one pins the height error…
ModelledLunar reference-station accuracy and coverage against published claimslunar_station (over lunar_dpnt,lunar_service)
Capability
The lunar-station-coverage kind run under the public assumptions of the published material on a lunar differential reference station (four navigation satellites in 24 h frozen elliptical orbits, 10 m 95 % SISE, 0.30 m code noise, 1 s and 600 s correction latency, a sub-decimetre station survey, the published example station site, the 300 km and 600 km south-polar circles) and scored against the sources' numeric claims under six one-sided bars pre-registered before the run: coverage of the whole south pole at sub-metre p95 (K1, K2), sub-metre and decimetre-level corrected p95 (K3, K4), sub-metre kept at 600 s latency (K5), and accuracy kept from 10 km to 500 km baseline (K6)
Oracle
Reference: the numeric claims of five public sources registered with URL, retrieval date, SHA-256 and licence in tests/fixtures/lunar_station/SOURCES.md (arXiv:2602.08432v3; IOAG Cislunar PNT Workshop 2025 and 2026 slides; UNOOSA ICG WG-B Lunar PNT, June 2024), transcribed with page references in tests/fixtures/lunar_station_published/CLAIMS.md. Bars, cases and inputs were pre-registered in tests/fixtures/lunar_station_published/PREREGISTRATION.md (revision 2, ebac1e14, committed alone before the run). 0.36 first run, a finding (stays MODELLED): 1 of 6 bars passes. Coverage is 0.00 on both circles against >= 0.95, and the corrected p95 3-D error is 280.6 m (300 km), and 293.0 m at 600 s latency, against <= 1.5 m. The decimetre bar (K4) could not pass at the sources' 0.30 m code noise, which reflects a conflict between the sources rather than a model result. K6 passes (ratio 1.39 against <= 1.5), but as a ratio of two geometry-dominated figures it is evidence about decorrelation in neither direction. Cause: the satellites' elements are not public, and on the ILLUSTRATIVE in-tree elements a user and the station share four satellites above 5 deg in only 8.2 % (300 km) and 6.3 % (600 km) of epochs, at near-singular geometry (uncorrected p95 3.8 km against the sources' < 10 m). The sources' simulations used a filter, a height constraint and three satellites in common view, which the snapshot kind does not model. The comparison therefore neither confirms nor refutes the published 'sub-metre over the whole south pole'. It shows that this model's result depends on orbits that are not published. Not a statement about any real system; not affiliated with or endorsed by any agency
Tests
tests/lunar_station_published_oracle.rs::recorded_finding_one_of_six_bars_passes (the recorded figures against the registered bars); tests/lunar_station_published_oracle.rs::published_claims_comparison_run (ignored…
The lunar protection level's geometry-to-covariance step — (GᵀG)⁻¹ over the all-in-view and every single-fault sub-geometry, projected onto the local vertical and the horizontal block trace — and its statistical kernel: the Bonferroni detector multiplier, the Gaussian upper tail of every fault hypothesis, and the root solve mapping an integrity-risk budget onto a protection level, over a 7-point lunar service volume
Oracle
Two established third-party implementations composed, both run offline, with the fixture regenerated end-to-end and verified byte-identical. RTKLIB 2.4.2-p13 (T. Takasu, BSD-2-Clause), commit 71db0ff, src/rtkcmn.c compiled from C source with -ULAPACK: lsq() → matinv() → ludcmp()/lubksb() inverts the normal matrix by Crout LU with partial pivoting, an algorithm unrelated to kshana's hand-written Gauss-Jordan invert4(); numpy.linalg.inv (LAPACK getrf/getri) re-inverts every one as a third independent inverse, agreeing with RTKLIB to 1.112e-12 relative. SciPy 1.17.0 / NumPy 2.4.1 (BSD-3-Clause) supply norm.isf, norm.sf (Cephes ndtri/ndtr) and optimize.brentq against kshana's Numerical-Recipes incomplete-gamma series, bisected inverse and 200-step bisection — and SciPy's direct sf is a different numerical route into the far tail than kshana's 1 − Φ, so the deep-tail cancellation is measured rather than mirrored. The vertical axis is the radial unit vector (the definition of local vertical on a spherical Moon) and the horizontal variance is the trace of the horizontal block, so no East/North azimuth convention is shared. EXPLICITLY NOT VALIDATED BY THIS ROW: σ_URE, the per-satellite fault prior and the illustrative Moonlight/LCNS-class constellation are MODELLED inputs handed to the oracle as given; and the FORM of the single-fault MHSS integrity equation is transcribed from the same published bound kshana implements — a shared closed form is a shared assumption, so this row covers a wrong evaluation of that equation, not a wrong choice of it. No third-party MHSS ARAIM implementation is reachable (RTKLIB exposes no protection-level entry point, no such PyPI distribution exists, Stanford MAAST is MATLAB), which is the outstanding upgrade
Tests
tests/lunar_protection_level_reference.rs (7 service-volume cases — south-polar, mid-latitude, equatorial, raised highland — 6 to 19 satellites, σ_URE 10/30/60 m, P_HMI 1e-3 down to 1e-7, protection levels spanning…
ValidatedARAIM MHSS protection levels against published reference vectorsaraim_reference (araim_reference_protection_levels,published_vectors),raim (araim_protection_level,araim_integrity_risk,normal_quantile)
Capability
The multiple-hypothesis solution-separation protection-level equation itself: all-in-view and per-fault-mode weighted least squares on an East/North/Up geometry with one clock state per constellation, the sub-solution sigmas, solution-separation sigmas and one-sided nominal-bias projections, the Bonferroni K_fa thresholds, the unmonitored-fault risk allocation, and the VPL/HPL that solve the integrity-risk equation. Runnable as the araim-reference-check scenario kind
Oracle
EU-U.S. Working Group C ARAIM Technical Subgroup, Reference Airborne Algorithm Description Document v3.1 (20 June 2019), Appendix D worked numerical example — published INPUTS (10-satellite 2-constellation geometry matrix, C_int/C_acc diagonals, b_nom, P_sat, P_const, LPV-200 constants) and published OUTPUTS (VPL 18.3 m, HPL 13.45 m, EMT 7.2998 m, σ_v_acc 1.3694 m, K_fa_3 5.1083, and six constellation-fault intermediates), retrieved 2026-09-20, source SHA-256 7f42934488c5c2261363439fabd48385e39526df34d514f395e22b6990b6bdb7. Matched at the reference's OWN stated tolerance TOL_PL = 5e-2 m: VPL 0.0074 m, HPL 0.0437 m; EMT and σ_v_acc to their printed precision. The same subgroup's Milestone 3 Report (2016) Annex A §A.IX states the same example but is internally inconsistent — a sign typo in row 3 of G (+0.7477 where the 2019 document prints −0.7477; with the plus, the document's own σ_v_acc comes out 0.8690 m against its published 1.47 m), and a K_fa_3 evaluated at 57 fault modes while the document states N_fault_max = 1, i.e. 12, so its EMT follows the 57-mode threshold and its VPL/HPL do not. No conforming implementation can reproduce all of its outputs at once. Its geometry intermediates reproduce exactly and its protection levels are recorded as measured discrepancies (VPL 0.0171 m, HPL 0.0842 m, EMT 0.4806 m), excluded from the acceptance figure; the tolerance was not widened to absorb them. SCOPE: N_fault_max = 1 only — multi-event fault subsets are refused rather than truncated, and exclusion, the chi-square consistency check and the double-counting re-allocation step stay MODELLED
Tests
tests/araim_reference_vectors.rs (2 published vectors + 6 negative controls, fixture tests/fixtures/araim_reference/wgc_araim_reference_vectors.txt: dropping the nominal bias misses the published VPL by 2.6457 m, a 10…
ModelledARAIM with simultaneous multi-satellite fault hypothesesaraim_reference (add_v42_subset_plan,add_v42_protection_levels_to_order); raim (araim_subset_raim,SubsetRule)
Capability
araim_reference::add_v42_subset_plan: the ADD v4.2 subset-determination rule (P_THRES, FC_THRES, monitored subsets fewest-removed first) generalised from pairs to simultaneous combinations of up to max_order satellite and constellation events, with the exact prior P_no fault prod q_e; add_v42_protection_levels_to_order runs the ADD v4.2 protection-level path on it (order 2 is the existing path exactly). raim::araim_subset_raim (opt-in): advanced RAIM over that hypothesis set, each monitored subset one (p_fault, T, b, sigma) mode in the integrity-risk sum, unobservable subsets moved to the unmonitored risk, the budget scaled by the ADD allocation of the unmonitored risk, and the subset of largest normalised separation reported when a fault is detected. Priors needing deeper subsets than max_order are refused, never truncated
Oracle
Internal consistency (InternalConsistency): the order-2 reduction to the MAAST-checked ADD v4.2 path, independent re-derivation of the priors, and a constructed two-satellite fault. No MAAST or RTKLIB run beyond pairs has been made, so the deeper hypothesis sets and the subset RAIM's protection levels are not externally checked. araim_subset_raim shares araim_dual_raim's simplifications (one sigma for integrity and accuracy, one clock, the plain normal tail): it is not ADD-conformant
Tests
tests/raim_multi_sv_subsets.rs (deeper_orders_change_nothing_when_pairs_suffice; triples_are_refused_at_order_two_and_planned_deeper: N_fault,max >= 3, the unmonitored risk within P_THRES, a subset removing three or…
ModelledRAIM beside the clock-holdover figures of meritclock_integrity (run,sky_at,IntegrityCfg); scenario (Scenario::integrity); report (RunResult::integrity_fom); api (clock dispatch)
Capability
The optional [integrity] section of the clock kind (clock_integrity::run): from a stated site and constellation presets (geometry::sky_geometry), at every integrity epoch while the GNSS timeline is not denied, snapshot RAIM (raim::snapshot_raim: parity chi-square test, slope protection levels) on the satellites above the mask with seeded Gaussian residuals of the stated sigma and an optional stated bias; a separate integrity_fom block with per-epoch verdicts, protection levels, and availability. Denied epochs are reported unavailable, never extrapolated; the holdover figures of merit are unchanged and, without the section, the output is byte-identical
Oracle
Internal consistency (InternalConsistency): the protection levels are those of snapshot_raim, which is validated against RTKLIB in its own row, recomputed on the same geometry; fault-free and injected-fault behaviour. The sky is a representative preset geometry (not a dated almanac), the residuals are synthetic and sigma, the alert limits and the probabilities are stated inputs; nothing here is a statement about a real receiver's availability
Tests
tests/clock_integrity.rs (the_block_is_separate_and_absent_by_default: holdover figures unchanged, 111 denied epochs unavailable with no protection level, refused with runs > 1…
The cislunar arc-length observability threshold in the SPATIAL six-state CR3BP rather than the planar four-state, across DRO, L2 halo and L2 near-rectilinear halo families, with measurement noise entering the Gramian and TWO criteria reported because there is no single honest one: the published rank criterion, which is provably invariant to a homoscedastic measurement sigma (whitening multiplies the observability matrix by a scalar, leaving the relative singular-value spectrum and hence rank, defect and condition unchanged), and an estimability criterion — the first arc at which the formal 1σ position uncertainty of the chief's initial state falls below a stated bound — which is the one that does move with noise. Each threshold carries the epoch-grid bracket it sits in; an unreached criterion is null with a reason. Planar mode remains the default and out-of-plane families are refused there rather than flattened into a state that cannot represent them
Oracle
An independent row-echelon rank by Gaussian elimination with partial pivoting confirms every six-state rank verdict — a different algorithm sharing no code with the eigen/SVD route under test. The 6×6 CR3BP state-transition matrix this composes carries its own ReferenceImpl oracle against SciPy variational integration, and the halo/NRHO initial conditions come from a corrector that reproduces the published L2 southern 9:2 Gateway orbit. The existing square-root-information-filter leg is explicitly NOT counted as corroboration: it consumes the same Jacobians and reduces to the same normal matrix, so it is a consistency check between two numerical machines, and the emitted extension label says so. NOT externally anchored, and ExternalDataset is therefore declined rather than borrowed from the STM's row: the threshold arc lengths themselves depend on the MODELLED constellation design, the epoch grid, and — measurably, by up to 40× — on the singular-value tolerance. The published planar 2.09 h is reproduced exactly at rel_tol = 1e-6 and is itself tolerance-dependent (10.25 h at 1e-4, 0.42 h at 1e-8), which is a property of the criterion rather than of the orbit. No public dataset publishes an arc-length observability threshold for a chosen cislunar constellation. MEASURED RESULT: the spatial six-state threshold is 22.17 h for the L2 NRHO and 40.42 h for the L2 halo, and DOES NOT EXIST for the planar-DRO family the published claim was derived on — rank 4 of 6 with datum defect 2 and two exactly-zero eigenvalues, a structural defect no arc length recovers
Tests
cislunar_observability::tests and observability_gramian::tests (the default document pinned bit-for-bit by FNV-1a of its JSON, summary and SVG, with explicit-default values asserted to be a no-op and the eight…
A least-squares bias-plus-rate fit over a trailing window plus the principal periodic terms — annual, semi-annual and the two principal zonal tides for UT1; Chandler, annual and semi-annual for the pole — extrapolated with the last in-window residual carried forward: the class IERS Bulletin A uses, in place of the persistence predictor the published horizon rested on. Scored the only honest way: a forecast for T+h built from rapid Bulletin A rows at or before T, measured against the LATER-PUBLISHED Bulletin B final at T+h, with persistence scored over the identical epoch set against the identical finals beside it. A target epoch with no published final is dropped rather than re-scored against the rapid column; a periodic term the window cannot constrain is reported as rejected with the cycles it actually spans; an incomplete window is refused rather than quietly shortened
Oracle
Three independent routes, none of them a published prediction-accuracy figure. (1) An analytic signal with known coefficients — the only way to exercise the periodic machinery, since no committed series is long enough to admit an annual term. (2) The textbook closed-form least-squares solution, different algebra from the matrix solve under test, on real rows. (3) The genuine archived Bulletin A prediction rows the real 2026 product publishes, compared per lead as an AGREEMENT statistic and explicitly not as an error: 0.256 ms at 1 day, 0.695 ms at 2 days, 1.252 ms at 3 days, drifting to 3.885 ms at 9 days. So this is Bulletin A's CLASS, close at short lead, not Bulletin A. The predicted-versus-final residuals are real measured quantities over real IERS rows, but their magnitude is checked only against the DIRECTION of the comparison, never against an IERS-published accuracy number — reading a real product is provenance, not an oracle. MEASURED: at day 1 the operational predictor gives 3.78 m of Moon-frame error against persistence's 11.39 m (3.01×), at day 2 10.23 m against 21.72 m (2.12×), at day 3 20.33 m against 30.57 m (1.50×) — and it is WORSE beyond three days (0.87× at 5 days, 0.43× at 10), which the report emits rather than showing only the horizons that flatter it. NOT reproduced: the autoregressive residual filter and the tabulated zonal-tide reduction, the 365-day operational window is unreachable with the committed data, and NO archived earlier vintage of the series exists in this repository — so the archived-vintage table reports no rows rather than scoring a synthesised one. 0.30 external comparison, a finding (stays MODELLED): scored against the same finals over 178 archived Bulletin A issues with the crate's default 15-day window, the predictor's UT1 MAE is 2.4x to 12.6x Bulletin A's at 1 to 10 days (bar 1.5x) and 5.32 ms at 10 days, outside the 0.36 to 3.13 ms range of the 2nd EOP PCC; the pole is 1.5x to 1.85x Bulletin A's (inside 1.5x only at 6 and 7 days); 19 of 21 pre-registered conditions fail. Correction to route (3): the 2026 product rows it was compared with carry the IERS I (measured) flag, not P, so those agreement figures are against measured rapid values, not Bulletin A predictions. 0.30 round 2, a least-squares plus autoregressive predictor with the zonal tides removed analytically, against 178 archived Bulletin A issues at the round-1 bar (pre-registered 65a3a1f7), a finding (stays MODELLED): 12 of 21 conditions hold; the UT1 MAE is 1.73 to 2.11x Bulletin A at 2 to 10 days (bar 1.5x) while the pole is within 1.27x at every lead. 0.30 revision: with the offline input now classified by IERS flag, the route-(3) agreement figures are against real Bulletin A predictions (0.256/0.695/1.252/3.885 ms at 1/2/3/9 days against rapid rows -> 0.133/0.421/0.839/2.786 ms)
Tests
frame_eop::tests, realtime_frame_eop::tests and tests/operational_eop_predictor_reference.rs (analytic-signal coefficient recovery to 1e-9 at a 365-day window; a bias-plus-rate fit equal to the closed-form…
ModelledLunar frame datum from an observing campaignlunar_frame_campaign (composing lunar_vlbi,lunar_vlbi_fim,fim,lunar,lunar_frame_realise,frames,cio,lunar_frame)
Capability
The seven-parameter Helmert datum propagated from a SIMULATED OBSERVING CAMPAIGN instead of recovered from an injected transform. Earth stations observe a sourced catalogue of lunar-surface beacons over an explicit schedule; the lunar-VLBI delay partials are accumulated into a beacon-coordinate Fisher information matrix and pushed through the Helmert design A = [I₃ | [p]ₓ | p] into H = AᵀM_bA, so the reported datum accuracy is a function of the observing programme — exactly linear in the delay sigma and monotone in the arc through the libration the report MEASURES. The datum defect is the subject rather than a footnote: rank, defect, condition number, spectrum, unobservable directions in the seven-parameter basis, the weakest direction even at full rank and each parameter's share of it are emitted on every run, and any parameter the campaign does not constrain is published as NULL with a status. Beacon-error correlation is MEASURED, not assumed — exactly zero with the stations held fixed, and printed with its cost when they are estimated. Runnable as the lunar-frame-campaign scenario kind
Oracle
Closed-form identities and an independent in-repo route, with no external oracle claimed. The Helmert design — the only new derivative in the module — is re-derived by central finite difference of lunar_frame_realise::apply_helmert, a path sharing no expression with the analytic form. The accumulation it composes is lunar-vlbi-fim's, whose Jacobian is finite-differenced against lunar_vlbi::vlbi_delay_s there, and whose linear-algebra kernel is separately externally anchored against numpy in tests/fim_observability_reference.rs — this row borrows neither status. The physics oracle is structural rather than numerical: a beacon delay partial is the near-field DIFFERENCE of two near-parallel unit vectors, so the worst-determined translation direction MUST be the body-fixed direction to Earth, and the test asserts the computed answer against the separately measured direction. MEASURED: campaign-derived translation sigma 6.3975 m against the injected-transform scenario's 0.3125 m recovery error (20.5× tighter), while rotation and scale run the OTHER way at 2.32× and 13.5× looser — the injected path assumes one isotropic sigma and so spreads its error uniformly across seven parameters, which the delay observable does not. Rank 7/7 but condition 3.8e5, with one direction 99.57 % pure translation-toward-Earth and forty times worse than any other. WHAT THIS ROW DOES NOT CLAIM: that the campaign figure is right in absolute terms. It establishes that the figure now DERIVES from a schedule, a geometry and an error model rather than from a planted answer. Stays ReferenceImpl/MODELLED: no published lunar-VLBI campaign-plus-datum-covariance pair exists, the station network and delay sigma are ILLUSTRATIVE inputs (the beacon catalogue is sourced, the campaign is not), observations are treated as independent while a real session's troposphere and clock are correlated between nearby scans, and the ephemeris, clocks, troposphere and Earth-orientation parameters are held FIXED. 0.30 external comparison, a finding (stays MODELLED): with every beacon Jacobian row rebuilt from NAIF SPICE light times (DE440, ITRF93, MOON_ME; pre-registered aad03856), the schedule, Helmert rank, datum sigmas (within 4.7e-3), condition, weakest direction (0.07 deg), libration sweep and arc ordering agree on six scenarios within the pre-registered 1 % / 0.2 / 0.05 deg. NumPy on the committed engine Jacobian agrees to 3e-14 where the stations are fixed. With the stations estimated it differs by 6.8e-6 against a pre-registered 1e-9, whose condition-number premise was wrong (Helmert condition 2.1e8; independent NumPy routes disagree by 3e-6; a contribution from the engine 1e-9 pseudo-inverse fallback is not ruled out), so the row is not promoted on that pre-registration. 0.36 re-registration at a condition-scaled bar (pre-registered 3855e840; t = max(1e-12, 10 kappa 2^-53) from the conditioning of the joint information and its Helmert reduction), a finding (stays MODELLED, NOT MET): step 0 rules the pseudo-inverse fallback OUT for the 6.8e-6 gap (no matrix upstream of the Helmert matrix falls back on the six original scenarios, tightest margin a factor 2.6), the SPICE leg still holds on all six, and NumPy agrees within the bar on six of eight scenarios (default, the two arcs, the 3x delay sigma, the collinear defect and the fresh stations-fixed run on 2026-12-01; worst 0.13 of its bar). It fails on stations_estimated (datum sigmas up to 6.8e-6, condition 9.1e-6, independence-discarded norm 3.5e-6, against t = 5.9e-7), and on the fresh date the south-pole beacon has no observation, so the beacon information is rank-deficient (pseudo-inverse fallback upstream of the Helmert matrix in both fresh scenarios) and the unchanged NumPy oracle cannot evaluate stations_estimated_fresh (singular matrix). Results: docs/design/evidence/mr05-lunar-datum-cond/RESULTS.md
Tests
lunar_frame_campaign::tests (21 lib tests: the Helmert design against a central finite difference of lunar_frame_realise::apply_helmert — the module the datum is FOR, which computes no derivative and so pins the [p]ₓ…
ModelledIndependent-estimator corroboration of the cislunar arc-length thresholdcislunar_arc_recovery,batch_ls
Capability
A batch least-squares estimator that RECOVERS the chief's initial state from simulated inter-satellite measurements over growing prefixes of the same epoch grid the rank-vs-arc table uses, with the measurement partials taken as central finite differences of the composed forward model. It consumes none of the machinery the threshold was measured with — no analytic Jacobian row, no variational state-transition matrix, no singular-value or eigen decomposition, no rank tolerance, no square-root information filter — and shares only the dynamics, the initial conditions, the scalar observable and the epoch grid, each named in the emitted document. Two criteria, both swept: noise-free recovery (the rank analogue) and Monte-Carlo estimability (the estimability analogue, measured against a known truth rather than predicted from a covariance). Runnable as the cislunar-arc-recovery scenario kind
Oracle
A separate estimator in this codebase on a different algorithm and a disjoint code path: nonlinear weighted Gauss-Newton with finite-difference partials, its verdict read as a measured state-recovery error in km and mm/s. It has no expression in common with the Gramian's rank-and-covariance route, and the separation is ENFORCED by a source-text guard rather than asserted. MEASURED on the published planar DRO grid: the ESTIMABILITY criterion is CORROBORATED — the measured Monte-Carlo boundary is 5.739130 h against the formal 5.739130 h (ratio 1.0000), and the measured RMS reproduces the formal 1σ over 21 arc lengths to a geometric-mean ratio of 0.9804. The RANK criterion is NOT corroborated as a recoverability boundary: the estimator recovers from 0.782609 h against the Gramian's 2.086957 h, ratio 0.375, and at 1.826087 h — the last prefix the rank read scores 3 of 4 — recovers to 1.96e-5 of the a-priori displacement. The recovery boundary is unmoved over three decades of its own bound while the rank threshold spans 'never' to 0.782609 h over four decades of rel_tol, and the two coincide exactly at rel_tol = 1e-8. That span was first recorded as reaching 0.260870 h; that figure was an ARTEFACT, produced by a rank read that counted 4 directions from 2 measurement rows below the f64 noise floor, and it is corrected here rather than left standing. The rank read is now bounded by Sylvester's inequality with the reason emitted, and the corrected span saturates at exactly the arc where the independent estimator recovers, which strengthens this row's cross-validation rather than weakening it: the published 1e-6 is a conservative singular-value CONVENTION, not a statement about recoverability. Same pattern spatially: NRHO 9.290323 h against 21.677419 h, halo 22.978723 h against 39.829787 h, with estimability agreeing exactly in both. The planar-DRO six-state recovers at no arc, independently reproducing that family's structural datum defect from an estimator told nothing about it. NOT externally anchored — the dynamics and initial conditions are shared with the analysis under test and no public dataset publishes such a threshold — so ExternalDataset is declined
Tests
cislunar_arc_recovery::tests and tests/cislunar_arc_recovery_reference.rs (both boundaries and their ratios pinned on the published grid; the measured error curve compared to the formal covariance row by row with a…
ModelledLunar service volume from real, retrieved constellation geometrylunar_ephemeris,lunar_service
Capability
Accepts a tabulated Moon-centred state ephemeris (the evaluation of an SPK/BSP kernel) or a published constellation definition behind ephemeris_path, runs the identical coverage / DOP / protection-level sweep against it, and emits the σ_URE ranging requirement it implies BESIDE the unchanged illustrative Keplerian and perturbed results with the difference as its own named quantity. Every figure carries a provenance class that distinguishes kernel-derived from published-element-derived from modelled, so the two can never be confused in a downstream quotation
Oracle
The GEOMETRY is external and hashed: three fixtures whose numbers come only from documents retrieved with URL, retrieval date and SHA-256, regenerable by committed generators that verify the upstream hash and ABORT rather than emit a number — the LANS interoperability-demonstration reference constellation (NASA NTRS 20250009447, SHA-256 d1b916be…), the LNCSS case studies (NAVIGATION 70(4) navi.613, CC BY, SHA-256 4e294687…), and a genuine flown-spacecraft ephemeris for LRO, Danuri, Chandrayaan-2 and CAPSTONE evaluated from JPL's own reconstructed kernels via Horizons. NO lunar-navigation constellation kernel exists publicly — Moonlight/LCNS, LCRNS and LNSS are not flying and NAIF publishes nothing for them — and none was invented. The DERIVED σ_URE requirement has NO external oracle (nobody publishes the ranging accuracy a 50 m lunar HPL demands over this service volume), so it is checked against its own algebraic identity and, independently, by re-running the whole sweep at the computed requirement and confirming availability flips there. MEASURED: the published 8-satellite design needs σ_URE 2.9044 m at 100 % coverage against the illustrative constellation's 0.3591 m at 37.85 % — an 8.09× revision of a published number under programme rule R4. The qualitative conclusion survives (LNIS-class 30 m still does not close a 50 m south-polar HPL) but the shortfall was overstated eightfold. The 5-satellite LANS demo yields ZERO protection-level samples — five satellites cannot give the six-in-view a single-fault hypothesis set needs — and the requirement field is ABSENT rather than fabricated; likewise for the four real spacecraft at 0 % coverage. InternalConsistency is the honest kind: the INPUT data is external and hashed, but the quantity this row is about is validated only against itself. 0.30 external comparison, a finding (stays MODELLED): against the LNCSS case-study statistics of navi.613 on 346 grid points, 4 of 12 availability statistics fall outside the pre-registered 2 percentage points (case A: -4.0, +22.5 and -3.6 pp; case B: -12.1 pp). Likely causes: elements read in the wrong frame, two-body Kepler over 15 days against the paper's unstated force model, and a grid not reproducible from the text. The PDOP half was not evaluable (LANS prints no numeric statistic); the comparison test is committed ignored. 0.30 round 2, a finding (stays MODELLED): against an Orekit 12.2 / Hipparchus 3.1 oracle computing the whole report (Orekit propagation, DOPComputer visibility and DOP, Hipparchus-composed MHSS protection levels, the requirement solved per sample; pre-registered 3c8e879a) on all five retrieved geometries and two grids, 155 of 160 values agree within the pre-registered bars (2 pp, 5 %, 1 %), including the sigma_URE requirement for LNCSS A and C (3.879072208 m on both sides for case A) and the absent requirement for LANS and the orbiters. LNCSS B on the dense grid fails on its worst-sample statistics (sigma_required 0.000923 against 0.001511 m, hpl_max 1.63e6 against 0.99e6 m) where near-singular geometries (PDOP about 1e6) amplify the metre-level position agreement; its p95 requirement agrees (1.6343 against 1.6357 m). The worst-sample requirement is ill-posed on degenerate geometry. Against navi.613: 4 of 12 still outside 2 pp (pinned)
Tests
lunar_ephemeris::tests (format and frame parsing; Lagrange interpolation exact at nodes and on a linear track off-node; ICRF elements take the IAU 2015 reduction and are byte-equal to an explicit icrf_to_iau_moon…
ModelledUnit and provenance declared for every reported quantityfield_schema
Capability
A machine-readable schema giving unit, provenance class and definition for every numeric field of every scenario report, plus a single global gate that runs every registered scenario kind and fails if an emitted numeric field lacks either. The provenance vocabulary is closed and each class carries an evidence tier, with two classes deliberately mapped to inherits-scenario-label and depends-on-input rather than being assigned a tier the class does not determine
Oracle
The emitted document itself: every numeric leaf must resolve to an entry in that document's own units block. There is no external unit registry to check a DECLARED unit against, so a declared unit is a reviewed assertion and not a verified one — the gate verifies COMPLETENESS and WELL-FORMEDNESS, not truth, and Validated would be wrong because nothing external confirms that m is the right unit for a field named _m. An independent name-suffix cross-check was run over the 1,434 fields described AT THAT TIME: of the 691 carrying a unit-bearing suffix, 40 disagree with the declared unit and all 40 are explained (per-second suffixes, minima, aperture-seconds, newton-metres against a nanometre-looking suffix), so no wrong unit surfaced. That cross-check has not been re-run since, and coverage has grown past it — the figure is left at what was actually measured rather than restated at the current total. At the time of that change (0.27.0), coverage moved from 7 of 56 kinds to 60 of 61, and from 389 of 1354 fields to 1,742 of 1,744 (docs/field-units-schema.json, which a staleness check keeps current). NO kind remains uncovered (ceiling 0, FUT-28, 0.36): the last was sweep-nd, two of whose columns are caller-keyed so their units are data; they are now described as mixed - see note and the unit of every position is read from the SWEPT pack's own units table into key_units and metric_units (or taken from the axis's declared unit, which must agree with the pack's when both exist; null when neither exists, stated rather than guessed). cislunar-observability was the one before it, and came off the list when its released document gained a units block: the additivity pin that had named units as a forbidden key now PROVES the addition is additive instead — strip the block back off and the document still hashes to the two constants frozen before it existed, so a released value that moved underneath the addition still fails. The work also surfaced twelve unit or documentation defects in existing code, reported rather than fixed under R1
Tests
tests/field_units_global.rs (all 61 registered kinds run, one document shape each: every numeric leaf must resolve to an entry in that document's own units block; a malformed entry counts as missing for EVERY kind…
ValidatedLunar frame datum from a REAL observing campaignlunar_llr,realdata::llr_crd (composing fim,cio,frames,ephem,lunar_frame,lunar_frame_campaign)
Capability
The seven-parameter Helmert datum covariance, with its two simulated inputs replaced by measured ones: the schedule is the ground-transmit epochs of 337 archived ILRS lunar laser-ranging normal points (2015-04-08 to 2015-06-27, Grasse MeO 7845 and Matera MLRO 7941, all five retroreflector arrays) and every observation weight is that point's own archived precision bin_rms/√n_raw — median 5.13 mm of one-way range, read out of the file. Station coordinates IERS ITRF2020, reflector coordinates JPL DE430 Table 7. Runnable as the lunar-llr-datum scenario kind. The validated claim (0.36) is narrow: on the analytic default path, the datum covariance computation agrees with an independent SPICE light-time and NumPy oracle on fresh measured normal points; it is not an LLR accuracy and not evidence that the analytic Moon is accurate
Oracle
Closed-form identities, an independent in-repo route and, since 0.36, an independent SPICE and NumPy library oracle on fresh measured data (below); no published lunar-LLR datum-covariance pair exists to check against, so no accuracy is claimed. The only new derivative — the range partial with respect to the reflector's body-fixed position — is re-derived by central finite difference of the two-way light time it differentiates, a path sharing no expression with the analytic form. The physics oracle is structural: the partial of a range IS twice the line of sight, so the body-fixed coordinate along the mean direction to Earth must be determined far better than the two plane-of-sky coordinates, which only the libration reaches — measured at 50.7× to 71.7× across the five arrays against a libration sweep the report measures independently, with the isotropic small-angle prediction and the transverse anisotropy that explains the gap both printed. WHAT CHANGED: the schedule, the observation count and every observation weight are now measured; 337 of 349 archived points are used and the 12 that are not are skipped and counted, because ITRF2020 carries no position for Apache Point. MEASURED: datum translation sigma 1.851746e-2 m against the SIMULATED campaign's 6.3975 m (345×), rotation 36×, scale 24× — a ratio between a laser-range network and a VLBI-delay network, so a finding rather than a validation. Reflector information rank 15/15 with inter-array coupling exactly 0; Helmert rank 7/7, condition 2.4e4. WHAT THIS ROW DOES NOT CLAIM: that the figure is right in absolute terms. A real LLR solution co-estimates the lunar orbit, physical librations, Earth orientation, station coordinates and tidal and relativistic parameters, and reports decimetres (DE430 Table 7's own 0.12-0.27 m) where this bound is far smaller; this is a Cramér-Rao bound for a stated reduced parameter set, not an accuracy. Outside the validated claim (modelled, not validated): the Moon-centre ephemeris and the IAU 2015 body orientation are modelled, and troposphere, tides, station eccentricity, polar motion, UT1-UTC, relativistic delay and station clocks are absent. Their combined size is PUBLISHED rather than argued — observed-minus-computed one-way range 156,494 m RMS over the 337 points — and their effect on the covariance is BOUNDED rather than argued: re-solving the entire datum with every partial tilted by 0.1° (twice the measured worst-epoch ephemeris tilt, sign alternating) moves the deliverable by 0.288 %. EXTERNAL CHECK (0.30 round 2, pre-registered ba48ff6d), a finding (stays MODELLED): SPICE two-way light time (DE440, ITRF93, DE440 lunar orientation) with finite-difference partials, and a numpy inverse, on the same 337 measured normal points and weights, agree on bookkeeping, both ranks, zero coupling, the residual RMS (2.1e-5) and 15 of 16 covariance figures. The z-translation sigma is 1.03 % high against the pre-registered 1 % bar. With the Moon centre from DE440 it falls to 0.049 % (6d333da, seen data), and on fresh 2019 normal points the kernel-Moon datum holds every bar (row "…, kernel Moon centre"), where the analytic Moon also stays inside the bars (worst 3.8e-3, information only), so that comparison does not separate the Moon models. FRESH-DATA RUN OF THE ANALYTIC PATH (0.36, pre-registered c6971d0 before any point was fetched): the unchanged SPICE and NumPy oracle on 151 used EDC CRD v2 normal points of 2025-04..06 (a year no test had used; files re-fetched and oracle re-run independently, see the fixture NOTICE) at the same bars: bookkeeping, ranks, per-array counts and zero coupling equal; sigmas and norms worst 1.11e-3 (1 % bar), condition 8.95e-4 (2 %), across/along ratios worst 1.79e-3 (2 %); SPICE O-C 10.79 m. VALIDATED on that evidence as a Library check of the covariance computation only; the analytic residual RMS on that slice is 222 km, so the bars do not discriminate Moon models (as on the kernel row). Results: docs/design/evidence/mr06-llr-analytic-fresh/RESULTS.md
Tests
tests/lunar_llr_real_data.rs (11 tests: all 15 committed CRD files byte-identical to their recorded digests; all 349 records accounted for, 337 used and 12 skipped for a named reason; every archived range inside the…
ModelledLunar reference-station geodesy: added retroreflector and VLBI beacon against the five-array datumlunar_llr (over lunar_frame_campaign::beacon_information,schedule_for and solve_datum_information,lunar_vlbi_fim)
Capability
Optional extra_reflectors and extra_beacons inputs on the lunar-llr-datum scenario add a station_geodesy block: the five-array Helmert information from the archived lunar laser ranging slice plus the information an added retroreflector (ranged at the archived normal-point epochs at which the ranging station is above its horizon, each with that normal point's archived weight) and an added VLBI beacon (observed by the lunar-frame-campaign network and schedule with the Earth stations fixed, through the lunar-vlbi-fim delay partials) at stated sites put on the seven-parameter datum, re-solved, with each Helmert parameter's 1-sigma before and after, their ratio, the before and after norms, and the extreme eigenvalues of the Fisher difference. Absent or empty, the report is unchanged
Oracle
Internal consistency only: each added site is a distinct point and a laser range or a fixed-station VLBI delay touches one point, so its information A_p^T M_p A_p is positive semi-definite and adds; the test checks the Fisher difference and every variance against that. The five-array result is the existing real-data row and is unchanged. The added reflector reuses epochs scheduled for the five existing arrays, an ILLUSTRATIVE schedule, and the added beacon's network, schedule and delay sigma are the ILLUSTRATIVE lunar-frame-campaign inputs; the site is an ILLUSTRATIVE south-polar position. A south-polar site sees the Earth low on its horizon or not at all, which the existing look-angle tests express; neither site is ranged continuously. The figures are Cramer-Rao bounds for a reduced parameter set, not a solved datum and not a statement about any real station
Tests
tests/lunar_station_geodesy.rs (empty_or_absent_extra_lists_are_bit_identical: empty lists and the baseline example give the default report and summary byte for byte; an_added_site_never_increases_any_variance: the…
ValidatedLunar frame datum from a REAL observing campaign, kernel Moon centrelunar_llr (llr_geometry_with,LunarLlrDatumScenario); ephem_provider (KernelEphemeris); naif_kernel
Capability
The lunar-llr-datum seven-parameter Helmert datum covariance from archived ILRS normal points and their measured weights, on the kernel path: the geocentric Moon centre is read from JPL DE440 by the engine's own kernel reader (planetary_kernel_path; lunar_llr::llr_geometry_with, ephem_provider::KernelEphemeris). Only the Moon centre differs from the analytic row; the IAU 2015 orientation, the absent polar motion and UT1, the catalogues, selection rules, weights and linear algebra are the same. The validated claim is only that this path's datum covariance agrees with an independent SPICE and NumPy oracle on fresh, unseen normal points. The comparison does not discriminate Moon models and does not check the Moon centre: on the same slice the analytic Moon also passes every bar, the pre-registered mutation (the kernel Moon read one hour late) was not detected, and a Moon centre about 4,650 km wrong also passes (see the oracle text). It is not evidence that the kernel Moon improves the datum. Runnable as the lunar-llr-datum scenario with planetary_kernel_path (and normal_points_dir for another slice)
Oracle
NAIF SPICE Toolkit N0067 (spiceypy 8.2.0, MIT) two-way light times on DE440, ITRF93 Earth orientation and the DE440 lunar orientation (MOON_ME) with 100 m finite-difference partials, and numpy 2.3.5 (BSD-3-Clause, LAPACK) information matrix and inverse: the unchanged generator of the analytic row's comparison (ba48ff6d), run on a fresh slice (EDC CRD normal points, 2019-04..06) pre-registered in adc4406e before any of it was fetched, at the analytic row's bars (1 % sigmas and norms, 2 % condition and ratios, exact bookkeeping, ranks and coupling) and its oracle self-check (SPICE O-C 10.86 m < 100 m). Every quantity holds, worst 7.6e-5. Mutations: the planned one-hour Moon delay is NOT detected (the covariance depends on the spread of the lines of sight, not their timing); a structural mutation chosen afterwards and disclosed as such, dropping the down-leg partial, fails every sigma. Finding: on this slice the analytic Moon passes too (worst 3.8e-3), so the bar does not discriminate Moon models and this validates the covariance on fresh data, not a kernel-Moon improvement. Integrator check at the fold (not pre-registered): the kernel Moon read relative to the Earth-Moon barycentre instead of the Earth (residual RMS 4,649 km) and the kernel epoch without TT minus UTC (69 s; residual 2.6 km) both still pass every bar, while dropping sqrt(n_raw) from the measured weights fails all 16 banded quantities; the comparison is sensitive to the weights, partials and linear algebra, not to the Moon centre. Both sides read DE440: this validates the computation on DE440, not DE440. Outside the claim: an LLR accuracy (a Cramér-Rao bound for a reduced parameter set)
Tests
tests/validate_llr_datum_kernel_moon_fresh.rs::llr_datum_kernel_moon_matches_spice_and_numpy_on_fresh_normal_points (447 unseen 2019-Q2 normal points; sigmas and norms within 4.0e-5 of 1e-2, condition 1.6e-5 of 2e-2…
ValidatedBuilt-in analytic lunar ephemeris — its STATED ACCURACY BOUND checked against real dataephem (moon_position),lunar_llr,realdata::llr_crd
Capability
ephem::moon_position, the low-precision geocentric lunar series the crate falls back on when no DE/SPK kernel is present, measured against real lunar laser ranging and against a published planetary ephemeris over the same span. This row validates the bound the module states about itself, NOT an accuracy: the series is confirmed to be no worse than it says, not confirmed to be good
Oracle
TWO independent external datasets that share nothing. (1) 337 archived ILRS lunar laser-ranging normal points (EUROLAS Data Center, DGFI-TUM, 2015-04-08 to 2015-06-27, five retroreflector arrays, committed with per-file SHA-256 and a re-download-and-verify script): observed-minus-computed one-way range 156,494 m RMS, with station coordinates from IERS ITRF2020 and reflector coordinates from JPL DE430 Table 7, so the residual is dominated by the Moon-centre series. (2) JPL Horizons geometric geocentric Moon states over the same span: 112,567 m RMS radial, 195,655 m RMS vector, worst epoch 360,324 m = 0.0536°. A range residual responds to the RADIAL part of an ephemeris error, so (1) and (2) are the same quantity reached two ways and agree to 1.39× — which is what licenses the lunar-llr-datum scenario to attribute its residual to the ephemeris. The engine's own module documentation claims ~0.3° / ~few·10² km; both measurements are inside it, the angular claim with a 5.6× margin. SCOPE, stated so no reader can mistake it: this validates the stated BOUND, not an accuracy figure. The series remains unsuitable for any application needing better than ~1e5 m, and that limitation is the row's content as much as the agreement is. Regenerable offline by the committed fetch-and-verify scripts
Tests
tests/lunar_llr_real_data.rs::the_analytic_moon_series_disagrees_with_jpl_by_the_same_amount (12 weekly geocentric states, worst-epoch angle asserted < 0.3° and RMS < 5e5 m against the module's own documented claim…
ModelledGate integrity — a green that means what it saysscripts/gate.sh,scripts/check-gate-receipt.sh,scripts/check-repeatability.sh,scripts/install-gate-hook.sh
Capability
Three structural guards over the verification process itself, each closing a class of defect that reached the canonical gate as an intermittent or spurious red. (1) Every constructed temp path in the Rust sources must carry a per-call unique component; a process id is rejected, because cargo runs the library tests as parallel threads of ONE process and the pid is shared by all of them. (2) Every byte-for-byte or hash pin must declare, next to itself or in its module header, what it covers and what it deliberately excludes — so a cross-cutting change can tell at a glance which pins are in scope and which have silently become repository-wide change detectors. (3) A green is claimable only from the FULL suite: scripts/gate.sh runs cargo test --all, captures the true exit code with no pipe in the way, and writes a receipt naming the commit, the integration-binary count, the test count and the duration; the pre-push check refuses a push whose commit no valid receipt names, on a clean tree. Repeatability is sampled separately by scripts/check-repeatability.sh, which runs the library suite N times and fails if the set of passing tests differs.
Oracle
Closed-form: each guard is run over a fixed known-bad snippet that must fire and a fixed known-good snippet that must stay silent, both literal constants the detector cannot influence. The two known-bad snippets are the F22 and F25 defects in their original shape. This is an INTERNAL oracle and the row is MODELLED accordingly: it proves the guards discriminate the defect from its fix, not that the classes they describe are exhaustive. Their stated blind spots — shared state that is not a path, a pin held in a short named constant, a rare race that three runs do not sample, and a receipt that anyone who can write the file can forge — are written out in the module documentation of tests/source_guards.rs and in each script's header rather than left to be discovered.
Tests
tests/source_guards.rs (9 tests: both guards over the whole source tree, plus mutation fixtures that grade each guard in both directions — the F22 pattern reported on both colliding paths, the shipped atomic-sequence…
Scalar-time solution-separation TPL over heterogeneous sources; PL driven by worst-exclusion subset noise + UTC(k) bias
Oracle
P2: numpy 2.3.5 linalg.lstsq and scipy 1.18.1 (norm.isf, norm.sf, brentq) recompute, by their own algorithms, the subset weighted least-squares estimators, the general separation covariance sqrt(Δ Σ Δᵀ), the biases and the protection level from them on 300 committed inputs (tests/fixtures/tpl_scalar_numpy_oracle); the PL agrees within 1e-9 relative (worst 1.2e-10) and the driving exclusion subset is identical, tolerance fixed before the comparison. The multipliers, priors and bias overbounds are inputs, and the structure of the PL equation (mode list, ir/2 split, bias projection) is a stated input, as for the Lunar ARAIM protection-level kernel row; the validated claim is the linear algebra and the PL on those inputs, not the physical magnitudes of a timing scenario. ARAIM lineage Blanch 2015 / Joerger 2014
Running-max envelope overbound-covers (zero-piercing) a real metrological series at every tested lag tau>=90d on a disjoint segment; short-tau (<=30d) is a disclosed Modelled boundary
Per-source bias overbound b = (U/k_cov)·Phi^-1(1-tail/2) + ageing, inflating a published Type-B expanded uncertainty to an allocated integrity tail; bridges to tpl_scalar as bias_s
Oracle
Independent numpy + stdlib-statistics reproduction of the overbound closed form (InternalConsistency). The BIPM Circular-T [UTC-UTC(USNO)] series is a CITED input only — it is the SAME series already Validated for the R4 holdover-coverage row and is NOT re-validated here; reproducing its values proves nothing about R2. Deep integrity tail is Modelled. 0.30 external comparison, a finding (stays MODELLED): on the BIPM per-laboratory UTC-UTC(k) files (102 laboratories, 105 024 rows with published uncertainties) the overbound b = 2.576 u at the 1e-2 tail is exceeded by 50.8 % of rows (Clopper-Pearson 95 %: 50.4 to 51.1 %) and by more than 1 % at 101 of 102 laboratories. The published u is the uncertainty of the BIPM's determination of UTC-UTC(k), not a bound on the offset, which routinely runs to many times u; the overbound as specified does not bound real UTC(k) traceability bias. 0.30 round 2, an overbound of the realised UTC-UTC(k) offsets trained on 2015-2022 and tested on 2023-2026 BIPM Circular T data (pre-registered a212810e, pooled exceedance at most 1e-2), a finding (stays MODELLED): the pooled exceedance is 0.0333 (694 of 20 826 rows over 82 laboratories; 0.0127 without the 430 rows that have no bound by rule). D9 round 3, a pooled hierarchical ageing bound (utck_bound), pre-registered 698f3842 and scored prospectively on Circular T issues from 465: not yet run (issue 465 was published on 2026-10-12 with about 285 test values; the pre-registered 1000-value rule needs issues to about 468, and the run needs the archived 2026-09-30 per-laboratory files, since BIPM has since rewritten 87 of the 113)
Tests
integrity::hetero_budget::tests + tests/hetero_budget_reference.rs; tests/hetero_budget_utc_k_oracle.rs::overbound_exceedance_on_real_utc_k_finding (data-gated: skips with a message when the BIPM series is absent)…
N×N bias cross-covariance with rho·b_i·b_j off-diagonals for sources sharing a UTC(k) realizer; PROVEN correlated_fused_bias >= independent_fused_bias (naive independent allocation under-bounds the fused bias -> optimistic -> unsafe)
Oracle
Independent numpy reproduction of the cross-covariance and the correlated-vs-independent fused bias inequality (InternalConsistency). Representative source set; Modelled scenario.
Hand-rolled Cholesky Omega=LL^T and forward-substitution whitening z=L^-1 r (Cov(z)=I when Cov(r)=Omega); Mahalanobis square z^T z = r^T Omega^-1 r
Oracle
P2: numpy 2.3.5 linalg.cholesky (LAPACK potrf), solve (gesv) and inv on 201 committed SPD inputs (tests/fixtures/gls_whitening_numpy_oracle) check the hand-rolled factor, the whitened residual and the whitening operator; the Mahalanobis square is checked against solve(Omega, r), which uses no Cholesky, so the identity z^T z = r^T Omega^-1 r is confirmed by a separate route. All within 1e-12 relative (worst 7.9e-16), tolerance fixed before the comparison. The largest condition number is 19.3, so ill-conditioned matrices are not covered. GLS/Aitken 1935 whitening is Cited.
Score statistic (1^T Omega^-1 r)^2/(1^T Omega^-1 1) ~ chi^2_1 for a common-mode shift that solution separation (contrasts orthogonal to 1) cannot see; a common-mode shift inflates it while pairwise separations stay flat
Oracle
Independent numpy reproduction of the statistic + property test that a common-mode injection inflates it (InternalConsistency). Modelled scenario. 0.30 round 2 (pre-registered 47c45647), blocked (stays MODELLED): numpy 2.3.5 LAPACK gesv reproduces the statistic value on 261 committed cases (within 3.5e-15 well-conditioned, 1.2e-9 at condition up to 6.1e7); the chi-square(1) null law is a cited theorem and the flat pairwise separations an identity with no engine output to compare, so the scoping awaits an owner decision
Undetectable common-mode ceiling: the largest fault magnitude along a direction d that escapes BOTH separation (alpha_ss = T_ss / whitened contrast norm of d) and the common-mode statistic (alpha_cm = sqrt(T_cm 1^T Omega^-1 1) / |1^T Omega^-1 d|), reported as min(alpha_ss, alpha_cm). Under a positive-definite Omega the two blind directions (d proportional to 1, and 1^T Omega^-1 d = 0) are mutually exclusive, so the ceiling is FINITE for every nonzero d; it is infinite only for a zero direction or a non-positive-definite Omega. A direction outside the modelled common axis is blind to the common-mode statistic only and is bounded by separation
Oracle
P2: numpy 2.3.5 linalg.solve (LAPACK gesv, LU with no Cholesky) recomputes alpha_ss (the threshold over the whitened contrast norm of d after removing its GLS common-mode coefficient), alpha_cm and their minimum on 256 committed (Omega, d, T_ss, T_cm) inputs (tests/fixtures/gls_outside_omega_bound_numpy_oracle), including pure common-mode directions, directions outside the modelled common axis, and the zero-direction and non-positive-definite cases that must give +inf; pre-registered (5adea293; amendment fe364dd1, before any comparison, restating the matrix form); tolerances fixed before the comparison. The capability text was corrected by owner decision (2026-10-02): it had said the bound is infinite for a direction outside the common axis, which contradicted the code and this comparison. That the idealised single-contrast detector lower-bounds a real subset detector undetectable magnitude stays Modelled
ValidatedLunar datum identifiability - decomposition linear algebralunar_identifiability
Capability
Schur-complement marginal of the {lunocenter-X, scale} block of the 7-parameter Helmert datum Fisher: the scalar degeneracy metric lambda_min(S), the origin-X CRLB, and the origin-scale correlation
Oracle
Independent SciPy/NumPy reference (scripts/gen_datum_identifiability_ref.py -> tests/fixtures/datum_identifiability/scipy_ref.csv): the Schur complement, its eigenvalues, and the 2x2 inverse computed by an independent library. Scoped to the decomposition LINEAR ALGEBRA; the lunar correlation/CRLB MAGNITUDES from real geometry remain MODELLED.
Tests
tests/lunar_datum_identifiability_reference.rs::decompose_matches_scipy_reference (16 synthetic 7x7 SPD matrices spanning well-conditioned to {0,3}-near-degenerate; Schur lambda_min / origin CRLB / |corr| vs…
ModelledLunar datum null-space classification (LLR rank-additivity + libration defect-lift)lunar_identifiability,lunar_datum
Capability
Classification of the internal-ranging datum problem: a single range observation contributes rank 1 (6-dimensional null space); the origin-X and scale pair is near-null and physical DE440 libration lifts the defect to zero while the pair stays near-degenerate
Oracle
Engine-reproduced geometric structure plus the closed-form 2x2 Schur identity; consistent in STRUCTURE with Sosnica et al. 2025 (arXiv:2510.15484), whose reported magnitudes are NOT reproduced here. The classification is structural; the correlation/CRLB magnitudes under real geometry are MODELLED (reflector coordinates and orientation held fixed). 0.30 external comparison, a finding (stays MODELLED): against Sosnica et al. 2025, the range-Fisher origin-scale correlation (-0.993 / -0.988) is within 0.05 of the published -0.97, but on the paper's own five-reflector geometry the equal-weight 7-parameter Helmert correlation is -0.78 (the published value rests on per-reflector errors the paper does not print), and the origin CRLB (0.6 to 1.2 mm at 3 mm range noise) is 96 to 98 % below the paper's 3.07 cm against a 20 % tolerance. Magnitudes not reproduced
ModelledMulti-technique lunar datum information (LLR + lunar VLBI + orbiter range)lunar_datum,lunar_identifiability
Capability
Datum-Jacobian rows for Earth-reflector LLR range, two-station lunar-VLBI differential delay, and orbiter-to-beacon range over the 7-parameter datum, combined into one consistently-preconditioned Fisher; demonstrates that off-radial (orbiter / depth-diverse) tracking breaks the origin-X to scale degeneracy where transverse VLBI helps only indirectly
Oracle
Analytic-vs-finite-difference cross-checks of each technique's partials (internal); the improvement DIRECTION is a geometric fact. Beacon locations, schedules and noise are MODELLED/representative (see tests/fixtures/llr_geometry/NOTICE.md).
Tests
lunar_datum::tests (LLR/VLBI/orbiter analytic partials vs finite difference); lunar_identifiability::tests (adding_an_offradial_technique_collapses_the_origin_scale_degeneracy; transverse-vs-radial via crlb_diag)
MOON_PA_DE440 to J2000 rotation at arbitrary epochs inside 2014-01-01 to 2030-12-31 TDB, by geodesic interpolation R0·Exp(f·Log(R0ᵀR1)) between the nodes of a committed 6 209-row daily series (the try_ functions return OrientationSpanError outside it, the infallible ones panic; it never clamps); embedded at compile time so there is no runtime filesystem I/O and the WASM build carries it
Oracle
NAIF SPICE Toolkit (CSPICE N0067 through spiceypy 8.2.0, run as a tool): pxform from MOON_PA_DE440 to J2000 evaluating JPL's binary PCK moon_pa_de440_200625.bpc directly, with NAIF's own frame kernel moon_de440_250416.tf, at 2 000 random off-node epochs in the embedded window. The rotation angle between Kshana's and SPICE's matrices is within 1.7e-5 rad (30 m at the 1737.4 km mean radius, fixed before the comparison); measured worst 1.7e-6 rad (3 m), RMS 7.8e-7 rad, and worst 1.75e-6 rad on 6 000 further epochs drawn by the verifier. The daily nodes were themselves extracted from the same kernel (scripts/gen_de440_moon_pa.py, fixture SHA-256 3076f81ef95d83f5efa240ed4c7ccb422f109407dde841fcf28d42dc63586eb7), so this comparison tests the interpolation between them, which is the claim. Found by this comparison and fixed: the earlier element-wise linear interpolation with Gram-Schmidt kept the axis but not the angle of the 13.2 deg/day rotation and was in error by up to 2.0e-4 rad (340 m) near the quarter points; the earlier figure of about 14 m came from a guard that samples interval midpoints only. The series covers 2014-01-01 to 2030-12-31 and errors outside it (round 2); its 2024-2025 nodes regenerate from the same kernel to 4.3e-13. The SPICE off-node comparison above covers 2024-2025 only, which is the validated window; the extended nodes come from the same kernel and script but are not compared off-node. A geometry substrate for identifiability analysis, not a sub-metre orientation product
ModelledLunar LLR datum geometry substrate (principal-axis reflectors, ground stations, analytic range partials)lunar_llr_geometry
Capability
The five near-side retroreflector arrays in PA body-frame metres and the LLR ground stations in geodetic coordinates; reflector body-to-inertial placement through the DE440 PA orientation; analytic partials of Earth-to-reflector range with respect to the 4-parameter datum; and the LLR-only Fisher matrix exhibiting the lunocentre-X to scale degeneracy that motivates the 7-parameter analysis
Oracle
The computation is checked internally: every analytic partial against a finite difference, and the zero-datum case against the nominal range. The CATALOGUE is checked externally and independently - the DE440 principal-axis coordinates (Park et al. 2021, tests/fixtures/llr_geometry/de440_retroreflectors_pa.csv) are cross-checked against Table 6 of the DE430 surface-coordinates memorandum (tests/fixtures/lunar_llr/de430_retroreflectors_pa.csv, machine-extracted from a SHA-256-verified PDF by tests/fixtures/lunar_llr/generate_de430_retroreflectors_pa.py) and agree to 1.12 m, while the mean-Earth realisation of the same five arrays differs by 672 to 871 m. The DEGENERACY STRUCTURE is checked against a published result: Sosnica et al. 2025, 'Definition and Realization of the International Lunar Reference Frame', arXiv:2510.15484, which reports for the lunar principal-axis frame a lunocentre-X to scale correlation coefficient of r = -0.97, and that LLR fails to reach the actual centre of mass of the Moon with an accuracy better than 12 cm because of that correlation. This module recovers the same structure (|corr(t_x, scale)| between 0.9 and 0.9999, datum defect <= 1) from an LLR-only Fisher design. The corroboration is across ephemerides, not a round trip: the paper combines INPOP21a, DE430 and EPM2021, whereas this module is driven by DE440 libration. What is NOT reproduced is magnitude - the correlation here is about -0.988 against their -0.97, and the 4-parameter CRLB is sub-millimetre against their 12 cm achieved floor, because orientation and reflector coordinates are held fixed here and solved there. Geometry and degeneracy structure only; no accuracy claim about a solved datum. 0.30 external comparison: the measured-range parts of this module (reflectors(), stations_itrf(), the DE440 PA placement and the barycentric light time llr_bcrs_one_way_m) are validated in their own row, LLR measured-range model, against 192 ILRS normal points (RMS 2.81 m against 10 m). This row stays MODELLED because the measured comparison does not exercise the rest of its claim. The geodetic station catalogue stations() rounds APOLLO to 0.001 deg; with a geocentric light time it gives 22.7 m on the same points (test reflector_ranges_against_ilrs_normal_points; 96 m with the earlier element-wise orientation interpolation). The analytic range partials and the LLR-only Fisher are computed through llr_range_m and reflector_inertial with the analytic Moon series (ephem::moon_position, about 0.3 deg from the GCRS) and no light time
The one-way Earth-station to reflector range of a lunar laser-ranging normal point: the five near-side retroreflector arrays in DE440 principal-axis (PA) body-frame metres (lunar_llr_geometry::reflectors); the stations of stations_itrf(), namely Grasse 7845 and Matera 7941 from ITRF2020 at epoch 2015.0 with their velocities, placed in the Geocentric Celestial Reference System (GCRS) by the IERS 2010 transform with Bulletin A polar motion (lunar_vlbi::station_inertial_position_itrs), and APOLLO 7045 at its operator-published approximate geocentric position (no velocity); the reflector body-to-inertial placement through the DE440 PA orientation (lunar_orientation::try_de440_moon_pa_body_to_inertial, which errors outside 2014-2030); and the IERS Conventions 2010 Section 11.2 barycentric light time (lunar_llr_geometry::llr_bcrs_one_way_m: the Eq. 11.19 body-to-BCRS transformation of station and reflector, and the Eq. 11.17 Shapiro delay from the Sun, the Earth and the Moon). INPUTS, not part of the claim: the barycentric Earth, Moon and Sun states and potentials (DE440 through SPICE), IERS UT1-UTC and polar motion, and the measured round trip converted from TT to TDB. HARNESS-SIDE, not part of the claim: the Mendes-Pavlis zenith delay with FCULa mapping, in the test, with elevation from the geodetic stations() catalogue. Not modelled: tides, station eccentricities, reflector thermal terms
Oracle
Measured: 192 ILRS CRD v2 lunar normal points (EUROLAS Data Center, DGFI-TUM, 2024-04 to 2024-06; Grasse 151, APOLLO 32, Matera 9; committed verbatim with SHA256SUMS). The predicted one-way range is compared with c times half the measured round trip in TDB. Bar fixed before the comparison: RMS at most 10 m over all 192 points AND over the 160 points of the two ITRF2020 stations alone. The bar is from the third pre-registration amendment 2b94bf6a, committed before the engine change and the run; the 10 m figure is unchanged since round 1. Measured: 2.81 m (mean 2.67 m) on all points; 2.96 m on the ITRF2020 stations; APOLLO 1.94 m; 2015 slice (349 points, secondary) 3.82 m. Mutation: zeroing the Sun Shapiro legs gives 10.19 m, which fails. History, disclosed: 96 m with element-wise orientation interpolation; 22.7 m after geodesic interpolation with the geodetic catalogue; 10.94 m with ITRF2020 stations and polar motion but a geocentric light time (a common +10.9 m offset). The barycentric model was written after that offset was seen. It is the published IERS 2010 Section 11.2 model with no free parameter. The third amendment replaced a self-imposed all-ITRF2020/SLRF2020 condition after the BLOCKED result. The ITRF2020-only RMS passes without APOLLO, which has no ITRF2020 or SLRF2020 solution, and its position is the operator stated-approximate one. A common +2.7 m remains, below the bar and not tuned. The bar is loose relative to the result: a missing 7.6 m solar term still gives 10.19 m. Terms below the bar are not resolved one by one: with polar motion zeroed the RMS is 6.15 m (ITRF2020 stations 5.98 m), and with the Eq. 11.19 transformation removed it is 2.96 m; both would still pass. Station velocities and the Earth and Moon Shapiro terms are smaller still. The comparison validates the assembled one-way range at 10 m RMS and the solar Shapiro term, not each listed component
Tests
tests/lunar_llr_geometry_range_oracle.rs::reflector_ranges_bcrs_iers2010_relativistic (192 ILRS normal points: RMS 2.81 m; the two ITRF2020 stations alone 2.96 m; both at most 10 m)…
ValidatedCoupled lunar frame and timescale gauge: the joint spatial-datum, clock-offset and clock-rate null spacelunar_gauge
Capability
assemble_coupled_info builds the 9-parameter coupled Fisher over [t_x, t_y, t_z, s, theta_x, theta_y, theta_z, d_tau, d_alpha]; classify_null_space returns a basis-invariant decomposition of its null space into spatial-only, temporal-only and genuinely coupled dimensions plus the spatial-to-temporal projector coupling norm; coupled_marginal_eigs gives the {scale, offset} Schur-complement marginal. On the committed networks the engine reports a full-rank datum (defect 0) for the well-posed case and an exact single rate-direction defect for the single-epoch case (elapsed time zero, so the rate is unobservable by construction), and computes the {s, d_tau} marginal in both.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: kshana computes the eigendecomposition with its own hand-written Jacobi solver (fim::sym_eig) and the marginal with its own pseudo-inverse; the oracle recomputes both with numpy/LAPACK (np.linalg.eigvalsh/eigh/pinv) via scripts/gen_coupled_gauge_ref.py. Agreement to rel<1e-3 is therefore an independent-library check of THIS CRATE ARITHMETIC against a standard implementation, which is the Validated claim and the whole of it. WHAT IS NOT: the assembly and the null-space classification are this module's own rules, and the oracle reproduces them from the same specification rather than from an independent source - that corroborates the implementation, not the rule. The beacon and orbiter geometry is Modelled and representative. The observation rows do carry the real DE440 Moon PA orientation (lunar_datum::orbiter_range_row_datum7 -> lunar_orientation::de440_moon_pa_body_to_inertial), but that is an INPUT both sides receive, not evidence about the answer, and it is recorded here as provenance rather than as the warrant for the status - the same distinction main already draws in the common-mode integrity row. Every row float is rounded to 6 dp before network.json is written, so both sides read identical IEEE-754 doubles and the only difference left is the solver.
Tests
tests/lunar_coupled_gauge_reference.rs::coupled_gauge_matches_numpy_on_real_de440_rows (9 eigenvalues, defect/dim_spatial/dim_temporal/coupled_dim/p_st_norm and the {scale,offset} Schur marginal for the well_posed…
ModelledRange to clock-offset degeneracy in one-way lunar ranging, and the geometric-diversity design lawlunar_gauge
Capability
The per-node radial-range to receiver-clock-offset degeneracy is exact at 1 m to 3.336 ns (one over c). The {scale, offset} Schur-complement marginal (coupled_marginal_fisher) shows that in a geometrically diverse network two-way ranging lifts the marginal by about the same factor as an equal count of additional one-way data, while in the idealised single-node regime the two-way lift is orders of magnitude larger. The design law that follows: geometric diversity, not two-way ranging as such, is what separates lunar frame scale from timescale offset in one-way ranging; a two-way or external time tie is required only in the low-diversity or single-user regime.
Oracle
Schur-complement marginal Fisher of the coupled information matrix, in closed-form linear algebra, checked against the two limiting regimes it predicts. The 1 m to 3.336 ns equivalence is the definition of the metre via c and is not an empirical result. The observation networks are representative Modelled geometry, so the design law is a statement about the geometry presented to it and carries no claim about any fielded lunar network.
ValidatedRelativistic clock-rate to frame coupling for the lunar timescalelunar_gauge,lunar_time
Capability
rate_frame_jacobian(_with) gives the sensitivity of the lunar timescale rate to the frame realisation, from the equatorial effective potential truncated at degree 2 (GM, equatorial radius, J2, rotation; LunarSurfacePotential): d_alpha/d_scale = L_m, about +3.1388e-11; d_alpha/d_velocity = -|v_moon|/c^2, about -1.11e-14 (time-dependent); d_alpha/d_r_radial = g_eff/c^2, about +1.806e-17 per metre. For a 1 m radial datum shift this is a rate perturbation of about 1.8e-17 (an arithmetic consequence of the radial entry), far below the roughly 1.7e-11 rate offset of a lunar timescale against TT
Oracle
Pre-registered (bb51058e): L_m against Ashby and Patla 2024 (AJ 167:149) Eq. (10), 3.13881(15)e-11, from the paper GM, a_m, omega_m and the J2 of the two degree-350 fields the paper cites (Bertone et al. 2021, AIUB-GRL350A/B), within 1e-4; the radial entry against pyshtools 4.14.1 (BSD-3) equatorial gravity of the same fields within 1e-4 (measured 8.3e-5 and 6.7e-5, a narrow margin); the velocity entry against JPL Horizons DE441 lunar speed within 4.5e-3. A degree-2 closed form of a degree-350 published value (residual degree 3-350; Reference, not P1). Disclosed: the entry-1 outcome was known before the pre-registration, and the shipped default radius 1738.14 km is the paper surface definition, adopted after round 1. The tidal contributions and the comparison band remain Modelled reference-surface figures
Tests
tests/lunar_rate_frame_coupling_preregistered.rs (entry1_d_alpha_d_scale_matches_the_published_l_m_for_both_cited_fields: L_m within 2.8e-6 of Ashby and Patla 2024 with J2 from AIUB-GRL350A/B…
ValidatedBasis-invariant classification of the coupled frame and timescale null spacelunar_gauge
Capability
classify_null_space decomposes the datum null space into spatial-only (dimension d minus rank U_T), temporal-only (d minus rank U_S) and genuinely coupled (rank U_S plus rank U_T minus d) parts, plus the spatial-to-temporal projector coupling norm p_st_norm, the Frobenius norm of the off-diagonal block of the null-space projector. Every one of these is a function of the PROJECTOR rather than of a chosen basis, so they are invariant under any orthonormal choice of null vectors - which is what makes the classification reportable at all.
Oracle
SciPy/NumPy (BSD-3-Clause; scipy.linalg.null_space and numpy.linalg.svd, LAPACK gesdd) classify the same committed matrices by subspace intersection, dim null([F; E_T]) and dim null([F; E_S]), and the SVD null-space projector, which differs from Kshana Jacobi eigenvectors and Gram-rank formulas; pre-registered (2519c87c): integers exact, p_st_norm to 1e-9 (P2, an independent numerical library), with the rank threshold fixed in advance and a spectral gap of at least 1e6 asserted. The row claim is the classification of a stated matrix, so P2 covers it in full. The constructed cases include direct-sum, coupled and mixed null spaces and random re-bases; the real-network dimensions of the coupled-gauge row are not re-counted here. The in-crate invariance test (a rotated null basis gives identical results) is kept
Tests
lunar_gauge::tests (classify_is_invariant_under_null_basis_rotation - an axis-aligned null basis and its 0.6 rad rotation yield identical defect, dim_spatial, dim_temporal, coupled_dim and p_st_norm…
ModelledRepresentative multi-technique measurement menu and the additive Fisher combination rulelunar_techniques
Capability
MeasurementBlock holds one candidate measurement campaign as a 7x7 Fisher information contribution plus a scalar relative cost; block_from_rows builds one from raw datum-Jacobian rows so every block is preconditioned identically through lunar_identifiability::assemble_multi_info; combine sums the information of a chosen subset, which is the correct rule because Fisher information is additive across independent measurements; representative_lunar_menu supplies a deterministic four-block menu (LLR, a transverse VLBI limb beacon, a near-side orbiter and a far-side orbiter). The budget-constrained design optimiser that consumes this menu is not part of this crate.
Oracle
Additivity of Fisher information across independent measurements is a closed-form property, checked here against an independently written element-wise sum and against the order-invariance and monotonicity that follow from it. NOTE ON EVIDENCE: the combiner previously had no test at all, and the one test in the module summed the two information matrices inline instead of calling it, so a broken combiner would have left the module green; the test now goes through combine and fails when it is mutated. All beacon locations, orbiter geometry, per-technique precisions and relative costs are representative choices rather than mission values (see tests/fixtures/llr_geometry/NOTICE.md), so every degeneracy metric and CRLB figure reached through this menu inherits the Modelled status of lunar_identifiability::decompose.
Tests
lunar_techniques::tests (combine_is_additive_order_independent_and_empty_is_zero - empty selection gives the zero matrix, a single selection is the identity, a pair equals an independently written element-wise sum, and…
rotation-only (6-parameter translation+rotation) decomposition of the DE440/INPOP21a/EPM2021 geocentric-Moon disagreement into a reducible common ICRF frame-tie and an irreducible lunar-orbit-orientation excess (the full 7-parameter Helmert fit is a separate public function NOT covered by this external check: its evidence is an internal analytic-recovery unit test plus tests/lunar_helmert_fit_cross_check.rs, which agrees it against the Validated exact Gauss-Newton fit in lunar_frame_realise and pins the opposite-rotation-sign convention between the two)
Oracle
Three independent authoritative ephemerides — DE440 (JPL), INPOP21a (IMCCE), EPM2021 (IAA RAS) — sampled via IMCCE calceph, cross-checked by an independent numpy SVD least-squares fit. WHY THIS IS Validated WHERE lunar_common_mode IS NOT, on the same three ephemerides: here the reported quantity IS a property of the published data — their mutual disagreement — so the ephemerides are the oracle, not an input. In lunar_common_mode the same disagreement feeds a MODELLED constellation geometry and the answer is a property of that model, which is why that row is correctly Modelled/ReferenceImpl.
Tests
tests/lunar_interop_budget_reference.rs (raw / rotation-residual / theta_moon / theta_frametie / theta_excess / reducible / irreducible for 3 provider pairs vs the independent SciPy SVD lstsq fit in…
ModelledCross-provider consistency tolerance for a mixed-provider userlunar_interop_budget
Capability
consistency_tolerance: inverts a user position budget to a per-parameter inter-provider Helmert agreement requirement (origin/scale/rotation) at a reference lever arm, optionally inflated by the P1 single-provider realization CRLB
Oracle
Worst-case triangle bound on the Helmert point-Jacobian action (closed form)
Tests
lunar_interop_budget::tests (monotonicity in budget; RSS budget reduction under a per-provider CRLB; worked rotation tolerance at the lunar lever arm; binding-term = rotation)
ModelledMulti-provider lunar interoperability error budget and frame-vs-ephemeris design lawlunar_interop_budget
Capability
interop_budget: reducible (common frame-tie) vs irreducible (dynamics) split under PerProvider / CommonFrameTie / CommonEphemeris conventions, with the irreducible-fraction design-law metric (~0.69 on real DE440/INPOP/EPM: a common frame tag alone leaves the dominant floor)
Oracle
RMS aggregation of the Validated per-pair splits under each convention (closed form)
Tests
lunar_interop_budget::tests (CommonEphemeris zeroes the budget; CommonFrameTie < PerProvider; irreducible_fraction > 0.5 on the real splits; convention ordering)
On the real-DE440 per-node lunar network (each row a differential inter-node one-way range built via pernode_range_row with a real DE440 Moon PA-frame line of sight), the engine computes the weighted parity projector P⊥ = I − G(GᵀWG)⁺GᵀW (the rank-deficient pseudo-inverse form that survives the 8-dim datum⊕timescale gauge), the T1 detectability residual ‖P⊥·b‖ (in-range faults annihilated, generic faults detectable), the Baarda MDB non-centrality quadratic form cᵀWP⊥c (the correct single-P⊥ weighted form, NOT the double-P⊥ form cᵀP⊥WP⊥c) with MDB = √(λ₀/cᵀWP⊥c), and the Byzantine block-spark detect/identify counts of the stacked effective peer signatures P⊥·B_T (column-rank, not merely nonzero)
Oracle
Independent numpy/scipy linear-algebra reproduction on real DE440 Moon PA-frame line-of-sight rows (P⊥ from the SVD of the whitened design W^{1/2}G; the MDB non-centrality from the classical Baarda (1968) parity-subspace squared norm; the block rank from a numpy SVD singular-value count) — a genuinely different numerical route than the crate's cyclic-Jacobi eigendecomposition of GᵀWG. Scope (per tests/fixtures/faultobs/NOTICE.md): this validates the LINEAR-ALGEBRA PIPELINE on these rows only. The node geometry is Modelled. It is NOT an operational integrity guarantee, a protection-level certification, or any RAIM/ARAIM availability claim, and NOT a Byzantine-fault-tolerance mission guarantee — the tolerated-fault COUNT for this network is Modelled, while the detect>f / identify>2f block-spark BOUND is Cited (Fawzi–Tabuada–Diggavi 2014; Shoukry–Tabuada 2016; block-wise spark of Donoho–Elad 2003) and proven-instantiation in code
Tests
tests/lunar_faultobs_reference.rs (faultobs_matches_numpy_scipy_on_real_de440_rows: the 84×84 P⊥ element-wise, trace(P⊥) = parity_dim = 52 and rank(G) = 32, detectability norms + flags, the MDB quadratic form cᵀWP⊥c +…
ModelledByzantine block-spark detect/identify bound on the autonomous per-node lunar network (T3)lunar_faultobs
Capability
byzantine_bound instantiates the secure-state-estimation coding / sparse-observability bound on the per-node network geometry: a Byzantine peer injects an arbitrary linear combination of the measurements it participates in, and the network DETECTS any coalition of ≤ f = block_spark − 1 peers and uniquely IDENTIFIES any coalition of ≤ ⌊(block_spark − 1)/2⌋, where block_spark is the smallest peer-coalition whose stacked effective signatures P⊥·B_T are column-rank-deficient. Peer detectability is FULL COLUMN RANK of P⊥·B_j (a nonzero-but-rank-deficient block is a nonzero attack that lands in range(G) and leaves no residual), not merely P⊥·B_j ≠ 0
Oracle
Geometric instantiation of the block-wise spark / secure-state-estimation bound (Fawzi–Tabuada–Diggavi 2014; Shoukry–Tabuada 2016; block-wise spark of Donoho–Elad 2003; Candès–Tao 2005) — NOT the Lamport–Shostak–Pease 3f+1 consensus threshold ('Byzantine' names the arbitrary/colluding fault MODEL only). The specific tolerated-fault count is a Modelled property of the representative network geometry, checked for self-consistency by hand-constructed analytic cases and the count formulas
ModelledAutonomous-holdover temporal-gauge floor for the self-referential lunar constellation (T4)lunar_faultobs
Capability
holdover_floor establishes that in an ensemble of autonomous nodes connected only by inter-node measurements the common clock rate (δα_j += 1 ∀j) is an ensemble-time free parameter lying in N(GᵀWG): no amount of additional inter-node ranging can observe it (rate_in_gauge = true, temporal_gauge_dim = 2 with the common offset). This is the genuine holdover floor — the temporal analog of the rigid-frame position gauge. Adding one external absolute-rate tie (a direct link to an off-network reference) expels the common rate from the null space (rate_in_gauge = false, temporal_gauge_dim = 1), proving the floor is a real gauge, not an oscillator-model artefact
Oracle
Ensemble-time free-parameter argument (Percival 1978; Lewandowski & Thomas 1991) instantiated on the representative real-DE440 per-node network: a relative-residual null test of the analytic common-rate generator against GᵀWG, with the tie-broken contrast distinguishing the genuine gauge from an oscillator-model artefact. Representative Modelled geometry, not a certified timing budget
Tests
lunar_faultobs::tests — holdover_floor_self_ref_and_anchored (self-referential net: the common rate lies in N(GᵀWG), temporal_gauge_dim 2; after one external rate-tie row: rate expelled from the gauge…
Three facets of the single decomposition ℝ^{state_dim} = N(GᵀWG) ⊕ range(GᵀWG) with the 8-dim datum⊕timescale gauge as N(G) and the state_dim − 8 = 32 observable directions as range(G): (T5) provider_mismatch_split classifies a common multi-provider bias (∈ range(G)) as UNDETECTABLE — needing an external tie — and a differential bias (∉ range(G)) as self-monitored / DETECTABLE; (§0) the unification is that the SAME range(G) blind subspace governs the undetectable fault class, so the datum gauge N(G) and the fault-blind subspace range(G) are one geometry, not two; (g7) slope returns the Brown protection-gap slope ‖Π_obs Δx̂‖ / ‖P⊥b‖_W → ∞ for a fault b ∈ range(G) (estimator corrupted with zero parity) and finite for a detectable fault
Oracle
Closed-form parity-projector / pseudo-inverse algebra on the representative real-DE440 per-node network: the common/differential split, the N(G) ⊕ range(G) unification and the Brown (1992) protection-gap slope all follow from range annihilation P⊥·G = 0. Representative Modelled geometry, not a certified protection level
Tests
lunar_faultobs::tests — provider_mismatch_split_common_undetectable_differential_detectable (common bias undetectable, differential detectable), slope_infinity_in_range_finite_detectable (g7: slope ∞ for b ∈ range(G)…
ValidatedTelecom-timing MTIE and TDEV on a holdover time-error seriestelecom_timing (mtie_sliding,mtie_curve_ns,tdev_curve_ns); allan (mtie,time_deviation)
Capability
The maximum time interval error and time deviation the telecom-timing scenario kind reports, on a telecom-scale holdover record: an O(n) monotonic-queue sliding-window MTIE (telecom_timing::mtie_sliding, needed so a day of one-second samples stays fast) and the engine's TDEV (allan::time_deviation), evaluated on the grid the kind reports and read back out of a full run that ingests the series
Oracle
allantools 2024.06 — an independent third-party frequency-stability library — mtie() and tdev() on tests/fixtures/telecom_timing/holdover_te_series.csv, the series parsed from the same 17-significant-figure text on both sides; the reference records the series' SHA-256 and is regenerable offline via tests/fixtures/telecom_timing/generate_telecom_timing_reference.py. Validates the ESTIMATORS on this record, not the synthetic holdover that produced it
Tests
tests/telecom_timing_reference.rs (17 MTIE and 12 TDEV averaging factors on a committed 2 048-sample chip-scale-atomic-clock holdover series with aging, flicker and temperature, vs allantools 2024.06 to <1e-12 / <1e-9…
ModelledITU-T telecom synchronisation masks with PASS/FAIL and margintelecom_timing (MASKS,eprtc_a_holdover_limit_ns,check_curve,check_mask)
Capability
Transcribed limits of ITU-T G.8272 (07/2025) PRTC-A/B, G.8272.1 (2024) Amd. 1 (07/2025) ePRTC locked and ePRTC-A holdover (including the Table 3 time-error envelope), G.8273.2 (2023) Amd. 2 (11/2025) T-BC/T-TSC classes A-D and G.8271.1 (2022) Amd. 3 (05/2025) reference point C, each open or closed interval as the table states it, with every 'for further study' entry left out; a verdict and a margin at the binding point per check, and the time to exceed each budget. Runnable as the telecom-timing scenario kind
Oracle
Transcription of the named Recommendations (freely downloadable from itu.int), each limit carrying its table or clause, listed with every unconfirmed or unimplemented item in docs/TELECOM-TIMING.md; the breakpoint-continuity identities the tables imply are checked in the tests. A standards transcription with internal consistency checks, not a conformance test and not an external validation. 0.30 round 2 (pre-registered 0f8c3f11), blocked (stays MODELLED): the ten mask values and thirteen ePRTC-A holdover values the Recommendations print as worked numbers (G.8272 Figs. 1-2, G.8272.1 Amd.1 cl. 8.2.1 and Figs. V.1-V.3, G.8271.1 Amd.3 Figs. 7-2 to 7-4) are reproduced inside half a printed unit, most of them table constants or breakpoints; G.8271.1 Fig. 7-2 labels 200 ns at 1.3 s against its own Table 7-1 (197.5 ns). The G.8273.2 class A-D limits have no printed worked value, and the G.8272.1 ePRTC locked breakpoint labels (on solid construction lines) were excluded by the pre-registered selection rule, so those remain transcriptions
Tests
telecom_timing::tests (mask_boundaries_follow_the_tables_exactly — open and closed ends, the τ = 400 s gap of G.8271.1 Table 7-3, continuity at every breakpoint the tables imply…
Seven constant-temperature MTIE and TDEV wander-generation masks for the telecom-timing kind: G.811 (09/1997) PRC (autonomous), G.812 (06/2004) Types I, II and III, G.8262 (10/2024) EEC Options 1 and 2 and G.8262.1 (11/2025) eEEC (locked mode), including the power-law pieces the tables write (8 τ^0.5 and the like), each interval open or closed as printed, measured through the 10 Hz filter the clauses state (taken as applied before sampling); the reference scenario scenarios/telecom-synce-eec.toml sets a free-running OCXO against the three EEC masks. Not implemented: the variable-temperature allowances, holdover, transients, jitter and the G.811 frequency accuracy
Oracle
Transcription of the named Recommendations (downloaded from itu.int on 2026-10-13), every row re-read against the text, with the breakpoint continuity the printed coefficients imply bounded in the tests (for example 40 · 10^0.4 = 100.48 ns against the 100 ns plateau of G.812 Table 4). A standards transcription with internal consistency checks, not a conformance test and not an external validation. G.812 (06/2004) is listed as superseded by a pre-published 08/2026 edition and G.811 has an Amd. 1 (04/2016); neither could be read, so neither is checked. The free-running OCXO scenario overstates the oscillator's contribution at observation intervals long against a loop's time constant; it says nothing about any equipment clock
ModelledClock-log formats (RINEX clock, TimeLab ASCII, TAPR TICC) read to a normalised phase seriesclock_log (read_input,MeasuredPhaseReport,MEASURED_UNITS,TELECOM_LOG_UNITS); telecom_timing (SeriesInput::clock_log); api (clock kind [clock_log])
Capability
clock_log::read_input: one parser layer that turns a RINEX clock file (AS/AR records of one named clock, versions 2 and 3.00 to 3.04), a TimeLab-style ASCII phase or frequency log, or a TAPR TICC timestamp or interval log into a uniform phase grid, with tau0 stated, from a header, or the median spacing, every gap, off-grid reading and duplicate counted, and a refuse, longest-run or linear gap policy; wired opt-in into the telecom-timing series ([series.clock_log], adding a clock_log block) and the clock kind ([clock_log], adding a measured block with ADEV, TDEV and MTIE, the table's SHA-256 and the phase SHA-256; runs = 1). Without the table both kinds are byte-identical
Oracle
Pre-registered in tests/fixtures/clock_log/PREREGISTRATION.md (committed alone, 05bb1d17, before the fixture existed); recorded in RESULTS.md. Three seeded synthetic logs, one per format (a TICC gap of three epochs, a RINEX second clock to ignore), read back by an independent reader written from docs/CLOCK-LOG-FORMATS.md by the same author: grid, gap and fill counts equal and the largest phase difference 0 s (bar 1e-13 s). allantools 2024.6 oadev, tdev and mtie on that phase agree within a relative 2.5e-14 (bar 1e-10) at every averaging factor it evaluates; it does not evaluate MTIE over the whole record, which the engine reports, so that one point per log has no oracle value and the bar as worded is not met there. MODELLED: synthetic logs, and the reader oracle shares the author; real TimeLab, TICC and RINEX clock files from instruments were not compared
Tests
clock_log::tests (hand-written logs per format: headers, separators, units, frequency integration, TICC channels and modes, gaps and policies, refusals); tests/clock_log_allantools.rs (the pre-registered comparison…
ModelledClock-log formats linuxptp (ptp4l, phc2sys, ts2phc) and chrony (tracking, statistics) read to a normalised phase seriesclock_log (parse_linuxptp,parse_chrony,grid_timed_offsets)
Capability
clock_log::read_input with the formats ptp4l, phc2sys, ts2phc, chrony-tracking and chrony-statistics: the logged offset of a linuxptp servo line (ns, the line's own bracketed time stamp, readings before the servo first reports s2 left out and counted) or of a chrony tracking or statistics log (s, by column position, UTC date and time) becomes a uniform phase grid through the same gap policies as the other formats; a source selector (clock) names one of several interfaces, clocks or addresses; the logs feed the telecom-timing series, the clock kind and the holdover-test kind
Oracle
Fixtures written for Kshana from the public line layouts of the daemons, with made-up values (NOTICE.md in the fixture folder); no log from a real installation was read and no oracle compares the parse with another reader, so a layout that differs between daemon versions can still be refused or misread. The estimators fed by the phase series are the ones validated against allantools in the rows above
ModelledHoldover test scored from an operator-supplied time-error recordholdover_test (HoldoverTestScenario,recovery,linear_fit); telecom_timing (check_mask,check_envelope,budget_results)
Capability
The holdover-test kind: a time-error record (a linuxptp, chrony, TICC, TimeLab or RINEX clock log, or time_s,time_error_ns pairs) is split at the two times the operator states for GNSS removal and restoration into a baseline, a holdover and a recovery step; the baseline is measured against the ITU-T masks written for a locked clock and the holdover against those written for holdover or an autonomous clock, the ePRTC-A holdover envelope and the time-error budgets; the holdover reports the least-squares frequency offset the system held, and the recovery whether and when |TE| came inside a stated band and stayed for a stated dwell, with the peak and the largest step. Every limit is the one the telecom-timing kind transcribes; the two times are the operator's statement and are not verified
Oracle
Hand computation on records built from formulas whose slope, step, extremum and recovery time are known before the run (the tests state them), and the masks, envelope and budgets of the telecom-timing rows above. No laboratory measurement of a real holdover was scored and no external oracle was compared, so the kind makes no claim about any clock or system; a timing-lab partner is needed before any per-clock-class figure could be promoted
ModelledGNSS signal availability in the Space Service Volume and beyond, compared with the ICG SSV bookletssv (constellations,geom,sweep,pattern,check)
Capability
The gnss-ssv kind's availability sweeps: per constellation (GPS, GLONASS, Galileo, BDS, QZSS, NavIC) and combined, the share of (user, epoch) pairs with at least one and at least four signals and the longest outage, for a user on a sphere (the booklet's 2,562-point global shell, 14 days at 1 minute) or on a trajectory (GEO, HEO, lunar transfer), from the Keplerian orbits of the booklet's annex C, a constant minimum radiated power inside the reference off-boresight angle, Earth blockage by the sphere plus 50 km and a C/N0 threshold; user antennas from the booklet's table B10
Oracle
The ICG Space Service Volume booklet (UN OOSA ST/SPACE/75, 2018), itself a simulation, compared under a pre-registration committed before the run (xval/gnss-ssv/PREREGISTRATION.md; results in xval/gnss-ssv/RESULTS.md, pinned by tests/ssv_booklet_results.rs, misses included). Run 1 on the pre-registered reading: the minimum-radiated-power constants and the lunar bar agree (S0, S6 met); the upper and lower SSV, RF-threshold, GEO and HEO bars are NOT MET, almost all through GLONASS, for which the booklet gives 20 and 28 degrees in tables B2 and B6 and 26 and 34 degrees in table 4.2. A diagnostic variant with table 4.2's angles agrees in every cell of the upper SSV, lower SSV and the six GEO missions (it does not replace run 1; RF, HEO and lunar cells still miss). The measured GPS L1 panel patterns are not reproduced by the parametric stand-in (RMS 6 to 9 dB against a 3 dB bar). Agreement shows the booklet's stated method is reproduced, not that GNSS signals behave so
ModelledLimb-ray ionosphere, plasmasphere and atmosphere effects on grazing GNSS signals, and the weak-signal receiver design, in the gnss-ssv kindssv (limb,weak)
Capability
A Chapman ionosphere, a Gallagher-type plasmasphere and a neutral atmosphere (refractivity bending, oxygen absorption, defocusing) integrated along the straight ray between transmitter and user: slant total electron content, its plasmaspheric part, excess delay, tangent altitude and the loss on grazing links; and a weak-signal receiver design (coherent and non-coherent integration, data-bit limit, aiding, Doppler search) that sets the C/N0 threshold from the acquisition and tracking-loop models
Oracle
Closed forms and published magnitudes in the tests (a Chapman layer's peak and the bending of the surface limb, a few TECU of plasmasphere on a GPS-to-GEO ray); the receiver design reuses the acquisition and tracking-loop models and their tests. No measured limb-sounding or weak-signal reception data is compared with it
ModelledOnboard GNSS navigation filter for high-orbit and lunar-transfer spacecraftssv::nav (run_nav,Ekf)
Capability
The gnss-ssv kind's [navigation] mode: a nine-state extended Kalman filter (position, velocity, receiver clock bias and drift, a solar-radiation-pressure scale) on pseudoranges whose availability, noise and C/N0 come from the same link geometry as the sweeps; truth is integrated with J2, Sun and Moon third-body and radiation pressure with shadow, the filter models J2 and the third bodies and estimates the radiation pressure; reports position and velocity errors inside and above the GNSS orbits, the longest measurement gap and the mean NEES
Oracle
Simulated truth against the filter's own estimate (a consistent filter's mean normalised estimation error squared is about 3 for the position) and determinism for a fixed seed. No flight or recorded high-orbit GNSS data is compared; the Sun and Moon are low-precision circular stand-ins
ModelledRelative navigation of a multi-spacecraft formation in a highly elliptical orbit from GNSSssv::formation (run_formation,RelEkf)
Capability
The gnss-ssv kind's [formation] mode: a chief and deputies each run the absolute filter; the deputy's position relative to the chief is estimated by subtracting the absolute estimates and by a relative filter on single-differenced pseudoranges of the signals both craft track, with the relative dynamics the difference of two nonlinear trajectories
Oracle
The same simulated truth as the absolute filter: the relative filter's error against the true relative state, and the check that it does at least as well as subtracting two independent absolute fixes. No flight formation data is compared
ModelledNavigation and timing error budget of a chain of mission phases from launch to entry (mission-thread)mission_thread (scenario,handover,conversions,budget,phase)
Capability
The mission-thread kind: a document of phases, each one of eight registered kinds, run in order with the 6x6 covariance, the clock and the frame handed on under a stated rule (carry; inflate, white acceleration noise over a stated gap on a straight-line model; reacquire, dead reckoning then the source floor, never below it, never a reset); frame conversions GCRF, MCI and MCMF and timescale conversions through TT, each its own budget row, with the Moon's state from an optional DE440 kernel cut and the analytic series as the stated fallback; the margin against a stated requirement, the weakest link, the events and a timeline figure. Every number is a stated input or a model of one
Oracle
Hand calculations in the tests (the straight-line covariance growth, the mean sidereal rotation, the secular lunar clock rate) and, for the Moon's state only, the committed DE440 kernel cuts against pinned JPL Horizons values (worst 2.6 m and 7e-6 m/s). No mission data is compared with the chain; its phases are rows of their own. The document is a stated chain with ILLUSTRATIVE inputs, public sources only and no agency endorsement
ModelledLaunch and ascent to a parking orbit with GNSS acquisition and fade (launch-ascent)mission_thread::earth::launch (LaunchAscent)
Capability
The launch-ascent phase: a constant-velocity position and velocity covariance with white acceleration noise (the dead-reckoning grade), a GNSS fix of stated accuracy applied after the acquisition time and outside stated fades, a bias-and-drift clock disciplined by the same fixes, and the parking-orbit insertion state
Oracle
The Riccati steady state and the hand calculation of the prediction in the tests. Nothing here models a launch vehicle or a receiver; every number is a stated input
ModelledRelative navigation to the docking envelope in low Earth orbit over sensor legs (leo-rendezvous)mission_thread::earth::rendezvous (LeoRendezvous); cw_dynamics; mission_thread::sensors
Capability
The leo-rendezvous phase: the relative state in the target's LVLH axes with the closed-form Clohessy-Wiltshire transition, a 6x6 Kalman covariance with white acceleration noise and position updates whose variance is the sensor's at the current range, over legs of a stated sensor (ILLUSTRATIVE optical camera, lidar and radio-ranging forms); the docking window is the time from which the relative sigma stays inside the capture envelope to the end of the phase. The chaser's absolute covariance is handed on unchanged and a row says so
Oracle
A one-axis hand calculation of the predict and update in the test; the Clohessy-Wiltshire transition is the one cw_dynamics checks against an independent numerical integration. The sensors are ILLUSTRATIVE and no real rendezvous is compared
ModelledTrans-lunar injection and cislunar cruise: propagated mean state and covariance with tracking updates (tli-and-cruise)mission_thread::earth::tli_cruise (TliCruise); propagator; ssv::sweep
Capability
The tli-and-cruise phase: the mean state by the engine's Cowell propagator (Earth point mass and J2, the Sun and Moon as third bodies, the Moon from the thread's DE440 kernel where it covers the segment and the Sun from the analytic series); the covariance by a forward-difference state-transition matrix of that propagation (the nominal run and six perturbed runs per segment) with white acceleration noise; the TLI burn's size and pointing error; gnss-ssv fixes whose availability and PDOP come from the space-service-volume sweep along an osculating orbit standing in for the arc; range and range-rate tracking along the geocentric direction scaled by the ground-network duty, with the Joseph-form update
Oracle
Liouville's theorem (the determinant of the state-transition matrix of the conservative part is one) and a finer difference in the tests; the mean state is compared with a public trajectory in the row below, which it misses. The burn error, the updates and the covariance are not compared with anything
The entry phase: the return coast of tli-and-cruise with no burn, ending at the entry interface; the flight-path angle and its sigma from the full covariance by a finite-difference gradient; the probability that the angle is inside a stated corridor; the peak deceleration at the nominal angle and at both corridor edges from the engine's planar ballistic point-mass entry (skip-out detected) beside the Allen-Eggers closed form; the landing footprint from a stated downrange sensitivity and the interface covariance across the orbit plane. No lift, guidance or heating is modelled
Oracle
Hand calculations in the tests (the angle sigma is the perpendicular velocity sigma over the speed; the Gaussian mass in a corridor). The planar entry is the engine's own and the Allen-Eggers form is a closed-form comparison; no flown entry is compared
ModelledGround-network coverage and tracking duty for a thread phase ([ground_network] of mission-thread)mission_thread::earth::ground_network
Capability
The ground-network overlay: stations with a minimum elevation, the visibility of the vehicle on a stated geometry (the entry state flown two-body with J2, the Moon's position for the cruise, or a stated duty), station handovers that lose a stated time each, and the resulting duty passed to the phase's ground-tracking update rate; the longest interval with no station in view is a budget row
Oracle
Zenith and antipode geometry checks and the expected visibility of a distant target in the tests. No real network, antenna or schedule is modelled
ModelledRelative navigation to docking in a lunar orbit (lunar-orbit-rendezvous)lunar_rendezvous (LunarRendezvousScenario); lunar_phase; mission_thread::lunar
Capability
The lunar-orbit-rendezvous kind and phase: a joint two-vehicle batch Gauss-Newton estimate over the arc (twenty-one parameters: both states and persistent signal-in-space position biases, common and independent), grown over a quarter, a half and the whole arc with step halving, from an absolute fix per vehicle (AFS pilot code-tracking bound, ILLUSTRATIVE relay set) and a crosslink range, under the crate's lunar force model; the mean normalised estimation error squared, the Clohessy-Wiltshire check on a circular orbit and the 6x6 covariance handed on
Oracle
Self-consistency under a pre-registration (xval/mission-thread-lunar/PREREGISTRATION.md, results and the estimator history in RESULTS.md): the mean NEES of the relative state is 5.85 to 5.93 against the band 5.2 to 6.8 in four configurations, MET on the fourth estimator design after three sequential-filter designs missed; the truth and the estimator share one force model, so this says nothing about model error. The propagator it shares is compared with LRO and CAPSTONE in a row below
ModelledLanding-dispersion budget at a lunar polar site (lunar-descent)lunar_landing (LunarLandingScenario); lunar_phase; terrain_nav
Capability
The lunar-descent kind and phase: a five-state per-axis inertial error model (position, velocity, accelerometer bias, tilt, gyro bias) over a powered descent, aided by AFS fixes (standalone or differential, with the clock-aided three-relay case), a terrain-relative fix (the engine's terrain-nav matcher on synthetic relief, or a Cramer-Rao bound on a LOLA DEM with an aperture-problem rule) and lunar time, with a Monte Carlo touchdown ellipse
Oracle
The closed-form inertial drift, the reused kinds' own results and the calibration of the 3-sigma ellipse (the fraction inside agrees with its expectation) in the tests. No flown or tested landing is compared
ModelledSurface positioning and timing availability by site and season at the lunar poles (lunar-surface)lunar_surface_map (LunarSurfaceMapScenario); lunar_phase
Capability
The lunar-surface kind and phase: relay availability and PDOP, Earth and Sun visibility and a disciplined surface clock for each site and season, with the terrain skyline of moonlight-service-volume; a single-site run equals that kind's availability and PDOP to the bit
Oracle
Exact equality with the existing kind for one site, and monotonicity checks, in the tests. The relay set is ILLUSTRATIVE and no surface measurement is compared
ModelledPowered ascent from a lunar polar site to an insertion orbit (lunar-ascent)lunar_ascent (LunarAscentScenario); lunar_phase
Capability
The lunar-ascent kind and phase: the five-state per-axis inertial error model run over a powered ascent with specific force along the path, aided by AFS fixes and lunar time, reporting the insertion position and velocity covariance carried to orbit and the engine-cutoff dispersion (reported, not carried)
Oracle
The closed-form inertial drift and monotonicity checks in the tests. No flown ascent is compared
ModelledThe lunar force model propagation against public reconstructed lunar-orbiter and cislunar trajectorieslunar_od (LunarForceModel); precise_od; mission_thread_lunar_oracles (example)
Capability
The propagator the lunar kinds share (GRAIL GL0660B degree and order 150, Earth and Sun third bodies, relativity, DE440 lunar orientation) compared with the public JPL Horizons trajectories of LRO (24 h, 10 minute samples) and CAPSTONE (one 6.5 day revolution, one hour samples), under bars fixed in a committed file before the run
Oracle
JPL Horizons spacecraft ephemerides for LRO and CAPSTONE, against bars pre-registered in xval/mission-thread-lunar/PREREGISTRATION.md (Amendments 1 to 6). LRO: MET, 33 m largest 3-D difference and 2.4 m RMS radial against 500 m and 60 m. CAPSTONE: NOT MET, a 136.8 km end-position difference against 25 km and a 28.4 km periapsis difference against 10 km, the causes not isolated. The row is Modelled because the CAPSTONE bar is missed and the comparison tests the shared propagator, not any filter, signal or budget
ModelledThe cislunar mean-state propagation against the public Artemis I trajectory (tli-and-cruise)mission_thread::earth::tli_cruise; propagator; mission_thread_oracle (example)
Capability
The mean-state propagation of the tli-and-cruise phase (Cowell, Earth point mass and J2, the Sun and Moon third bodies, the Moon from a DE440 cut) compared with the public JPL Horizons trajectory of Artemis I over a 72 h coast arc from 2022-11-16 18:00 TDB, at 12 hour epochs, under bars fixed in a committed file before the run
Oracle
JPL Horizons target -1023 (the NASA navigation as-flown post-TLI trajectory), against bars pre-registered in xval/mission-thread/PREREGISTRATION.md (Amendment 1 changes only the request span). NOT MET and asserted as missed: 45.3 km largest position difference against 5 km, and 0.249 m/s velocity difference at 72 h against 0.05 m/s. A diagnostic written after the verdict places the miss in five impulsive velocity steps of 0.09 to 0.22 m/s in the reference (the pre-registered reading that the window held no burn was wrong); over the other 67 hours the model reproduces the next hourly state to a median of 0.38 m. Tests the mean state of one coast arc only: no covariance, filter, update or other phase
inertial::cai_error_terms (library functions, not yet reachable from a scenario kind): the Raman-phase weighting function of a square-pulse three-pulse interferometer (Cheinet et al. 2008, Eq. 8) with the white-phase closed form (pi/2)^2 S_phi0/tau_R and a numeric route for any spectrum; the laser frequency noise the retro-reflection delay does not cancel (Le Gouet et al. 2007, Eqs. 9 to 11); the pure-defocus wavefront bias sigma_v^2/R of a thermal cloud and its reduction by a detection radius; the wrong-fringe-index probability erfc(pi/(sqrt(2) s)) for a classical-accelerometer aid and for two interrogation times, and its inverse
Oracle
Each closed form against an independent numerical route in the tests (frequency-domain quadrature of the exact transfer function and delay filter, a Monte Carlo over atoms for the wavefront bias, a Monte Carlo over shots for the fringe index) and its limiting cases. One published worked value: Cheinet et al. 2008 (arXiv:physics/0510197, Sec. IV) state that at tau_R = 20 us a white Raman phase noise below -111 dB rad^2/Hz is needed for 1 mrad per shot; the white term alone gives 0.99 mrad (the paper derives the level with its Eq. 17, flicker plus white). Internal consistency and one worked value, not an external validation; higher-order aberrations and any specific laser's noise are not modelled
ValidatedGreenhall and Riley exact equivalent degrees of freedom of the Allan, modified Allan and Hadamard variancesallan (edf_greenhall,greenhall_edf_core); assurance::uncertainty (edf)
Capability
allan::edf_greenhall (with the edf_greenhall_adev, _mdev and _hdev wrappers): the full algorithm of Greenhall and Riley, Proc. 35th PTTI (2003), Section 3 (Eqns 7 to 13 and the large-J coefficients of Tables 1 to 3), for difference orders 1 to 3, noise exponents +2 to -4 wherever alpha + 2d > 1, overlapping and non-overlapping, modified and unmodified estimators; NaN where it does not apply, including the unmodified white-PM case with ceil(M/S) <= d, whose closed form is not implemented. assurance::uncertainty::edf now shares the code and keeps its outputs bit for bit
Oracle
allantools 2024.6 ci.edf_greenhall, a third-party implementation of the published algorithm, run by scripts/gen_greenhall_edf_ref.py over the grid pre-registered in xval/greenhall-edf/PREREGISTRATION.md (committed alone, 696f68e6, before any comparison): at all 4 876 points where it returns a finite edf, kshana agrees within a relative 6.0e-13 (bar 1e-9); at its 46 NotImplementedError points kshana returns NaN; every case label occurs; the wrappers equal the core bit for bit; uncertainty::edf is bit-identical over 274 280 values (pinned as a SHA-256 before the refactor). At 32 points (alpha -4, d 3, unmodified, case 2.2) allantools fails in its Table 2 lookup, so they have no oracle value and are not counted. Stable32 was not run. 0.36 revision (docs/revisions/STUDIO15-greenhall-edf-portable.md): on portable maths the kernel takes its logarithm from a double-double evaluation rounded once (allan::greenhall_ln), and the unchanged Bar 1 holds again at all 4 876 points, worst relative 6.0e-13; the uncertainty::edf guard is re-based to the 0.36 portable dump (3 746 of the 274 280 values moved, at most 1.6e-12 relative, from the host-maths base)
ModelledOscillator holdover presets from public datasheetstelecom_timing (PRESETS,fit_noise,with_flicker_floor,synthesize_holdover); models (ClockModel)
Capability
Four presets — OCXO (Microchip OX-208), rubidium (Microchip 8040C), caesium (Microchip 5071A high-performance tube) and CSAC (Microchip SA.45s) — each carrying its datasheet's Allan-deviation maxima, aging and temperature bound; a non-negative least-squares white + flicker + random-walk frequency-noise fit scaled to envelope every datasheet point, a flicker floor never below the longest-τ figure, linear aging and a sinusoidal temperature term, synthesised as a seeded holdover after a GNSS loss
Oracle
The four named Microchip datasheets (document numbers and URLs in each preset and in docs/TELECOM-TIMING.md) for the input figures; the closed-form aging ramp and the fit's envelope property as internal checks. How the figures become a noise model — the fit, the floor rule, the 30-day month, the linear temperature reading — is a modelling choice; no preset is a measurement of a unit. 0.30 round 2, one-sided bars against measured units (pre-registered 43d19980), a finding (stays MODELLED): the IGS station THU2 Symmetricom 8040C in ESA MGEX final clocks has an Allan deviation 0.35 to 0.87 of the rubidium preset but a holdover error 1.07 to 1.60 of it; the 5071A record sits outside the caesium preset to 1000 s (measurement white phase noise, 30x at 1 s) and in holdover (1.15 to 8.74); the worst 2011 production SA.45s is 1.26x the CSAC preset at 1 s; the OX-208 has no measured record (blocked)
Tests
telecom_timing::tests (every_preset_carries_a_named_datasheet_and_figures; the_noise_fit_envelopes_every_datasheet_point; pure_aging_matches_its_closed_form — the discrete aging ramp within 0.1 % of (1/2)·D·t² over a…
ValidatedHoldover prediction from a measured clock record, checked on held-out dataslot_timing (ClockNoise::from_phase_record,fit_weighted,slot_budget,breach_after_sync_s)
Capability
The slot-timing kind's measured-record path: the overlapping Allan deviation of a phase record fitted by edf-weighted non-negative least squares in the white-phase and IEEE Std 1139 frequency-modulation basis (slot_timing::fit_weighted), then inverted for the coast time at which the time error reaches a threshold (slot_timing::slot_budget). The red-noise floor is measured from the record instead of assumed
Oracle
A real free-running atomic clock: 556 990 one-second phase samples of a 5071A caesium standard measured against a hydrogen maser (A. Wallin, distributed with allantools; pinned commit and SHA-256 in scripts/fetch_cs5071a.sh), the same record tests/cs5071a_reference.rs uses for the estimators. The held-out two thirds are an external measurement the prediction never saw. Data-gated: the realdata-clock workflow fetches the record and runs the test with KSHANA_REQUIRE_REALDATA=1. Validates the fit-then-invert path on a white-FM-dominated atomic standard over coasts up to about 50 000 s; not the datasheet or class sources, not the deterministic terms. NOT a crystal oscillator: on a measured OCXO record (tests/slot_timing_ocxo_holdout.rs, allantools, 5.5 h against a hydrogen maser) the held-out prediction is conservative, at 0.59 to 0.70 of the measured breach, and mostly outside the bar, because that oscillator's noise floor halved during the record; in sample it is within 0.98 to 1.17. The crystal case stays MODELLED. NOT atomic clocks in orbit either: on 10 GPS Block IIF satellite clocks (tests/slot_timing_igs_holdout.rs, IGS final 30 s clocks, 14 days, protocol written before download) 8 land within the bar and 2 (G25, G30) are optimistic by up to about 2x, so under the protocol the orbital case stays MODELLED
Tests
tests/slot_timing_cs5071a_holdout.rs (holdover_inversion_predicts_the_held_out_caesium_record: fit on the first third of the record, predict the one-sigma breach at 0.5, 0.75, 1, 1.5, 2 and 2.5 ns, measure it on the…
ModelledSeconds until a free-running clock leaves a time-indexed slot's guard, and the fix cadence that keeps it insideslot_timing (predicted_error_s,breach_after_sync_s,terms_at,slot_budget,ClockNoise::from_datasheet,SlotTimingScenario); clock_state (ClockClass::Tcxo,Ocxo,Rafs,source)
Capability
The slot-timing kind: the predicted error E(t) = k·√(σ₀² + σ_x²(t)) + (|y₀| + |c_T·ΔT|)·t + ½|D|t² inverted for the breach, the time left, the resynchronisation interval net of the fix latency and fixes per day, each term itemised with the dominant one named; clocks from the six ClockClass defaults (TCXO, OCXO and RAFS added, each citing one datasheet), the telecom-timing presets, an inline datasheet or a measured record; a breach beyond the source's longest averaging time flagged as extrapolated
Oracle
Closed forms for each term alone and the independent holdover::holdover_seconds root-find; the datasheet figures are the named documents (docs/SLOT-TIMING.md). A class default rests on an assumed red-noise floor and a datasheet source is an envelope of maxima, so neither is a measurement of any unit
Tests
tests/slot_timing_igs_holdout.rs (10 GPS Block IIF satellite clocks, 14 days of IGS final clocks: 8 within the 1.5 bar, G25 and G30 optimistic by up to about 2x, pinned); tests/slot_timing_ocxo_holdout.rs (a measured…
ModelledTiming protection level for a receiver in orbit under GNSS spoofingorbital_timing (orbital_timing,visibility_half_angle_rad,max_pass_s,crosslink_neighbour_is_outside_footprint)
Capability
orbital_timing::orbital_timing, reachable as the slot-timing kind's spoofing section: circular-orbit speed and period, the longest time a ground spoofer can reach the satellite per pass, the clock-aided monitor floor and CUSUM detection latency, the conditional protection level (floor plus coast over the latency), the pull a spoofer at a stated maximum ramp rate accumulates before the satellite leaves its footprint or an independent check runs, and whether each ground-contact or crosslink check is independent of one ground spoofer
Oracle
Reduction to the existing timing protection level and the circular-orbit and spherical-Earth visibility closed forms. Conditional on detection, one terrestrial spoofer at a stated ramp rate, an overhead pass on a non-rotating Earth; not validated on a spoofed receiver in orbit
Tests
orbital_timing::tests (reduces_to_the_terrestrial_tpl — exactly tpl::timing_protection_level_ns when the clock has no flicker, white-phase or random-run term; leo_geometry_matches_its_closed_forms…
ValidatedClosed-form L-band signal power spectral densities and spectral separation coefficientsnavsignal (Modulation::psd,spectral_separation_coeff_offset,q_from_ssc); spectrum (psd_nulls_hz,psd_peak_hz,main_lobe_null_to_null_hz,Jammer::ssc)
Capability
Unit-area power spectral densities of GPS L1 C/A and L2C (BPSK(1)), GPS L5 and Galileo E5a (BPSK(10)), sine-BOC(1,1) and Galileo E1 MBOC(6,1,1/11) (navsignal::Modulation::psd, with an MBOC variant added), their numerically located nulls and maxima (spectrum::psd_nulls_hz, psd_peak_hz, main_lobe_null_to_null_hz), and the spectral separation coefficient of a signal against any spectrum at any offset (navsignal::spectral_separation_coeff_offset) or against a tone, flat noise, a chirp or matched noise (spectrum::Jammer::ssc), with the anti-jam coefficient Q = 1/(R_c kappa)
Oracle
Published textbook values: the BPSK(n) main lobe of 2n x 1.023 MHz null to null and the anti-jam coefficients Q = 1 for a narrowband (CW) jammer and Q = 1.5 for a spread-spectrum jammer matched to C/A (Kaplan & Hegarty, Understanding GPS/GNSS, 3rd ed., section 9.4); the BOC(m,n) main lobes centred at plus or minus m x 1.023 MHz (Betz, Binary Offset Carrier Modulations for Radionavigation, NAVIGATION 48(4), 2001); the spectral separation coefficients -61.8, -64.8 and -67.8 dB/Hz for C/A with C/A, BOC(1,1) with BOC(1,1) and C/A with BOC(1,1) (Betz 2001; Hein et al., MBOC: The New Optimized Spreading Modulation Recommended for Galileo L1 OS and GPS L1C, Inside GNSS, May/June 2006), reproduced here from their Parseval autocorrelation closed forms. The BOC(1,1) maximum is not at 1.023 MHz: the lobe spans the carrier null to 2.046 MHz and peaks at 0.759 MHz, and the test pins both. The MBOC mix is the ICD definition, checked for unit area and linearity only
Tests
spectrum::tests (bpsk_main_lobe_null_to_null_is_two_n_times_1_023_mhz — BPSK(1) 2.046 MHz and BPSK(10) 20.46 MHz located numerically on the closed form; boc11_lobes_are_centred_at_plus_minus_1_023_mhz — carrier null…
ModelledL-band spectrum waterfall with per-band J/S and effective C/N0 under a scripted jammer timelinespectrum (SpectrumScenario,SpectrumModel::bin_psd_w_per_hz,SpectrumModel::band_state,effective_cn0_multi_dbhz,system_temp_k,Jammer::power_fraction_in,Jammer::duty)
Capability
The spectrum kind: a frequency-by-time grid of the L-band power spectral density (thermal floor k T_sys with T_sys = T_ant + 290 K (F - 1), the signals at their interface-specification minimum received powers, and continuous-wave, narrowband, chirp and matched-noise jammers with on/off times), each cell averaged over its bin and row (chirps exactly over whole and partial sweeps, jammers by duty), per-band effective C/N0 = [1/(C/N0) + sum (J/S) kappa]^-1 per row, J/S per band, in-band J/S, and an SVG waterfall with C/N0 bars. The report carries a cross-check against the jamming kind's chain on the same link inputs
Oracle
Reduction to the existing jamming kind's anti-jam equation and link budget (the same code, called on the same inputs), and the k T0 F noise-floor closed form. The signal spectra underneath are the validated row above; the jammer powers, timeline and front-end bandwidths are scenario inputs, the spectra are continuous (no spreading-code lines), and no automatic gain control, blanking or antenna pattern acts on the jammer. No measured jammed spectrum is in the repository to check the composite against. The jamming kind's Q table (CW and narrowband 1, broadband and swept 2 since 0.36: the textbook values, pinned to the navsignal closed forms) agrees with this model's CW value and differs from what it derives for noise matched to C/A (1.5) and for flat null-to-null noise without in-band normalisation (2.2); the report prints both
Tests
spectrum::tests (agrees_with_the_jamming_kind_chain — J/S equal to jamming::j_over_s_db and effective C/N0 equal to jamming::effective_cn0_dbhz with Q = 1/(R_c kappa) to 1e-9 dB, and the 32.105 dB anchor of the jamming…
ValidatedSigMF recording input and output, and Welch spectral estimates of complex IQsigmf (read,write,encode,decode,parse_meta,meta_to_json); spectrum (welch_psd,fft_in_place,synthesise_iq)
Capability
sigmf: read and write Signal Metadata Format recordings (JSON .sigmf-meta with the core global, captures and annotations fields; raw .sigmf-data as cf32_le, ci16_le or ci8, the integer decoders shared with realdata::iqif::load_iq), all on strings and byte buffers. spectrum::welch_psd: Hann-windowed, overlapped, averaged periodograms, density-scaled, on an in-crate radix-2 transform (spectrum::fft_in_place). Those two are the validated claim. spectrum::synthesise_iq draws the model as IQ, and the spectrum kind's [iq] section runs model to IQ to SigMF to Welch and compares with the model; its [recording] section estimates a real recording (native builds); the synthesis and those model comparisons are outside the validated claim
Oracle
scipy 1.18.1 scipy.signal.welch (BSD-3-Clause) with the same periodic Hann window, overlap, density scaling, no detrend and two-sided mean average: every bin within 1e-12 relative (observed 6.0e-14) over 27 cases, with segment counts taken from scipy's own spectrogram. sigmf-python 1.13.0 (LGPL-3.0, run as a tool) writes five third-party recordings (cf32_le, ci16_le, ci8, two captures with annotations, a non-zero first sample_start) that the crate decodes bit-identically with identical core fields, and reads the crate's recordings back bit-identically; the SigMF v1.2.6 metadata schema reports zero errors on the crate's metadata. Tolerances fixed before the first comparison. Outside the claim: spectrum::synthesise_iq and the [iq] and [recording] comparisons with the model (a synthesis has no single right answer; a periodic chirp shows lines, Fresnel ripple and edge tails the smooth model omits, total power within 2 %), the integer-to-float scaling convention (the crate divides by 32767 and 127, sigmf-python's autoscale by 2^15 and 2^7, which the specification leaves open; the comparison uses raw codes), multi-channel recordings and the data types the reader refuses
ValidatedPlanet positions across the solar system from the JPL Standish Keplerian elementsephem (standish_state,standish_elements,standish_nominal_error,ecliptic_to_icrf); ephem_provider (AnalyticSolarSystem); solar_system (SolarSystemScenario)
Capability
Heliocentric positions of Mercury, Venus, the Earth-Moon barycentre, the Earth, Mars, Jupiter and Saturn from Standish Table 1 (1800 AD to 2050 AD), and of all eight planets from Tables 2a/2b (3000 BC to 3000 AD, with the b, c, s, f mean-anomaly terms), following the JPL page's algorithm step for step and rotated to the ICRF (International Celestial Reference Frame) by the page's obliquity; the Earth is split from the barycentre by the Montenbruck & Gill lunar series and the mass ratio. Composed as ephem_provider::AnalyticSolarSystem so any body is available relative to any other, and runnable as the solar-system kind
Oracle
JPL Horizons DE441 geometric heliocentric vectors (ecliptic and ICRF), fetched after the pre-registration commits e5423015, 59dac98b and 7b218df5. Bar: per planet and component, RMS at most 1.0x the approximate errors of the Explanatory Supplement 3rd ed. Table 8.10.1 (the EMB row for the Earth); the Earth-EMB split is checked directly against DE441 399 minus 3, at most 0.3 deg and 8.1 km at every epoch 1950-2050 (the 8.1 km length bar comes from the lunar-series accuracy stated in Kshana own ephemeris documentation, citing Montenbruck and Gill, and was not checked against the book; the measured 6.09 km sits close to it). The RMS statistic was chosen knowing that an earlier 27-epoch comparison reached 1.87x nominal at its maximum; single-epoch errors reach 3.83x (3.87x in the ICRF case). The Supplement does not define its figures as RMS. Case B rotates both sides with the IAU 1976 obliquity typed in the test, so the rotation is not checked against itself. The default table selection of AnalyticSolarSystem is not exercised (the tests force Table 1). Uranus and Neptune from Table 1 are a separate MODELLED row
Tests
tests/solar_system_standish_preregistered.rs::standish_rms_errors_are_within_explanatory_supplement_table_8_10_1 (barycentres 1800-2050 from Table 1 and 3000 BC-3000 AD from Table 2, 8908 and 21857 epochs, worst RMS…
ValidatedSun, Moon, Mercury and Venus positions from the DE440 kernel at UTC epochs (KernelEphemeris)ephem_provider (KernelEphemeris::relative_position_utc,utc_to_tdb_jd2); naif_kernel; timescales
Capability
Geometric positions of the Sun, Mercury, Venus, the Earth and the Moon relative to one another (J2000/ICRF axes, metres) from a JPL DE440 kernel through ephem_provider::KernelEphemeris at UTC epochs: body names mapped to NAIF codes, UTC carried through the leap-second table and TT to TDB with the two-term TDB-TT series, the kernel read by the engine's own reader. A body the kernel holds only as a system barycentre (Mars to Neptune, Pluto in de440s) is refused, never substituted. The kernel-free Standish row stays the browser and fallback path
Oracle
Skyfield 1.54 (MIT), with its own leap-second table, its own TDB-TT and its own reader of the same NAIF de440s.bsp (jplephem), run as a separate program at UTC epochs. Pre-registered (7c23d520) before the fixture existed, the grid seeded from that commit's hash, with a bar of 5e-5 s times the relative speed (the engine's stated two-term TDB-TT accuracy) plus 1e-13 of the barycentric distance plus 1e-5 m, per component. Result: all 600 inside, worst 0.66 of the bar (2.96 m on a fast pair, the TDB-TT series' own error). Mutation: TT used as TDB fails at 33 times the bar. Both sides read DE440: this validates the provider and its time chain, not DE440. Outside the claim: the planets beyond Venus as bodies (de440s carries their barycentres only), the planetary moons, and precision below the two-term TDB-TT series (a few metres for the fastest pairs)
Tests
tests/kernel_ephemeris_skyfield_oracle.rs::kernel_ephemeris_matches_skyfield_at_utc_epochs (600 positions at 200 UTC epochs 1973-2026 drawn from the pre-registration hash; worst 0.66 of the bar, 2.96 m)…
ValidatedLight time between solar-system bodiesradiometric (light_time_solution,two_way_range,shapiro_delay); solar_system (link,link_on); ephem_provider (AnalyticSolarSystem)
Capability
The Newtonian one-way light time from any body to any other, received at an epoch, solved by the radiometric fixed-point light-time solver (radiometric::light_time_solution, transmitter at its retarded position) through solar_system::link_on on any EphemerisProvider; the solar-system kind runs it on the analytic solar system in the heliocentric frame (solar_system::link). The two-way range from radiometric::two_way_range and the Sun Shapiro delay from radiometric::shapiro_delay are reported separately and are not externally compared. Emitted for every body against an observer and for any extra link by the solar-system kind, and for the body-to-Earth downlink by the body-pnt kind
Oracle
JPL Horizons DE441 one-way down-leg Newtonian light time LT (VEC_CORR=LT, observer the Earth centre), pre-registered (876f8edc, amendment 35d22bad changing only the interpolation node spacing after the first fetch failed its 10 m interpolation precondition, before any light time was compared) with a bar of 1e-6 s per light time, the solver fed DE441 barycentric positions so the retarded-transmitter iteration is visible (a first-iterate solver misses by up to 0.1 s). On the analytic Standish positions the light time is only as good as those positions (see the planet-position row); the shipped link treats the Sun as an inertial centre, which stays inside that position error but is not separately validated. two_way_range and shapiro_delay are not externally compared
Tests
tests/solar_system_light_time_solver_preregistered.rs::light_time_solver_on_de441_positions_matches_horizons_lt_within_1e_6_s (Mercury, Venus, Mars, Jupiter and Saturn barycentres, the Sun and the Moon to the Earth…
ModelledUranus and Neptune from Standish Table 1, Pluto, and planetary velocitiesephem (standish_state,StandishElements::state_at_mean_anomaly,Planet::Pluto); solar_system (position_label)
Capability
Positions of Uranus and Neptune from Standish Table 1 (the default inside 1800 AD to 2050 AD), Pluto from the 1992 Table 1 row the current JPL page no longer lists, and the heliocentric velocity of every planet: the two-body derivative with the mean motion the elements imply plus the turning of the ellipse by the perihelion, inclination and node rates
Oracle
Measured against JPL Horizons DE441 (the same committed fixtures), but not validated: Table 1's stated figures for Uranus and Neptune are exceeded by up to 2.0 and 5.2 times against DE441 heliocentric positions (Neptune meets them against barycentric positions, which suggests the fit was not heliocentric), the page states no error for Pluto, and no published bound exists for the velocities. Tables 2a/2b meet their own figures for Uranus and Neptune and are covered by the validated row
Tests
tests/solar_system_horizons_reference.rs (standish_table1_uranus_and_neptune_exceed_the_nominal_error: pinned at 2.04 and 5.16 times the nominal longitude error, under six times…
ModelledPositions of the Moon and seven major moons (Phobos, Deimos, the Galilean moons, Titan)ephem (satellite_state,Satellite,SatelliteMethod,moon_icrf,moon_icrf_velocity); ephem_provider (AnalyticSolarSystem)
Capability
The geocentric Moon from the Montenbruck & Gill series (whose -1.3972 deg per century term already refers it to the J2000 equinox), and planetocentric positions and velocities of Phobos, Deimos, Io, Europa, Ganymede and Callisto from the JPL Solar System Dynamics mean elements in each moon's Laplace plane with the mean-longitude rate taken from the IAU synchronous rotation rate (the table prints the period to too few digits to hold the phase), and of Titan from the IAU rotation model (the moon on minus the body-fixed x axis at the mean distance), because Titan's printed Laplace-plane row does not reproduce Horizons even at its own epoch
Oracle
Measured against JPL Horizons satellite ephemerides (DE441, MAR099, JUP365, SAT441) committed in tests/fixtures/solar_system/horizons_satellites_icrf.csv with the query and retrieval date; no published accuracy exists for mean elements or for the rotation-model placement, so the pinned bars catch regressions and do not validate. Mean elements omit the resonant and solar perturbations, and Phobos' secular acceleration is not modelled
Tests
tests/solar_system_horizons_reference.rs (moons_track_horizons_to_the_measured_angles: worst angular error seen from the planet over 2000 to 2040, Phobos 5.65 deg, Deimos 0.38, Io 0.40, Europa 1.63, Ganymede 0.22…
ModelledPhysical constants of every solar-system body and the whole-system reportbody (Body::by_name,Body::facts,Body::prime_meridian,SOLAR_SYSTEM,BodyFacts); solar_system (SolarSystemScenario,position_label)
Capability
Body gains Mercury, Venus, Jupiter, Saturn, Uranus, Neptune, Pluto, Phobos, Deimos, Io, Europa, Ganymede, Callisto and Titan (gravitational parameter, reference radius, J2 where published with the radius it is referenced to, IAU pole and prime meridian), a name lookup and a BodyFacts record (NAIF code, class, parent, equatorial and mean radius); the solar-system kind reports them for all eighteen bodies with the positions, light times and an orbit track over one revolution, each body labelled VALIDATED or MODELLED
Oracle
Transcription of published constants, each cited where it is defined: gravitational parameters from the JPL Horizons body records and the JPL satellite physical-parameter table, radii from the JPL physical-parameter tables (IAU WGCCRE 2015), poles and prime meridians from the IAU Working Group on Cartographic Coordinates and Rotational Elements (mean values, without the T-rate and periodic terms), J2 from Smith et al. 2012, Iess et al. 2018 and 2019, and Jacobson 2009 and 2014. Checked for internal consistency (the Jupiter rotation period against the System III rate, each J2 against its reference radius), not against an external oracle. 0.30 round 2: the GM values were adopted from gm_de440 and the Phobos and Deimos elements from pck00011, so 0 of 84 constants of the fourteen added bodies now differ from NAIF (was 12), which is a transcription check and is not counted. 0.36: the IAU body-fixed orientation and the GM of the fourteen added bodies are split out to their own VALIDATED row ("IAU body orientation and gravitational parameter of the fourteen added solar-system bodies"); this row keeps what is left. J2 and the mean radius are cited, not validated; the positions in the report rest on their own rows; and the report prime_meridian_deg for the Earth, the Moon and Mars stays on conventional constants (up to 6.6e-3, 2.8e-2 and 9.6e-4 rad from IAU_<BODY>). 0.30 revisions: Phobos GM 7.087e5 -> 7.087546066894452e5, Deimos 9.62e4 -> 9.615569648120313e4, Uranus 5.7939506103e15 -> 5.793951256527211e15, Neptune 6.83509997e15 -> 6.835103145462294e15, Pluto 8.69326e11 -> 8.696138177608748e11 m³/s²; Phobos equatorial radius 13.1 -> 13.0 km; prime_meridian_deg now includes the quadratic and periodic terms
ValidatedIAU body orientation and gravitational parameter of the fourteen added solar-system bodiesbody (Body::iau_rotation_et,Body::iau_angles_et,Body)
Capability
Body::iau_rotation_et, the rotation from J2000 to the IAU body-fixed frame of Mercury, Venus, Jupiter, Saturn, Uranus, Neptune, Pluto, Phobos, Deimos, Io, Europa, Ganymede, Callisto and Titan with every century rate, quadratic term and periodic term the body carries in pck00011, and the gravitational parameter field mu of the same fourteen bodies. The Sun, Earth, Moon and Mars carry conventional constants owned by other rows and are not covered
Oracle
NAIF SPICE Toolkit (CSPICE N0067 through spiceypy) pxform('J2000', 'IAU_<BODY>') on pck00011.tpc, an independent implementation of the same IAU rotation model, and NAIF gm_de440.tpc for GM, the fixtures generated by a separate program that calls no Kshana code (tests/fixtures/body_orientation_spice_oracle/, NOTICE.md). Pre-registered f3fa98b (with the original 1e-9 rad bar of b4e0ff5a) before the fresh-epoch fixture existed; the fresh fixture was regenerated independently and is byte-identical. GM is a transcription of DE440 checked against the published file, not an independent determination of GM. Results: docs/design/evidence/mr08-body-orientation/RESULTS.md
Tests
tests/body_orientation_spice_oracle.rs::orientation_matches_spice_pxform_for_the_fourteen_added_bodies (41 epochs 1955 to 2044, 574 matrices within 1e-9 rad, worst Phobos 1.43e-10 rad)…
ModelledPositioning around any solar-system body with a local constellation and a deep-space link from Earthbody_pnt (BodyPntScenario,formal_sigma,fix_error); orbit (dop); batch_ls (gauss_newton); mars_pnt (chord_clears_sphere); mars_frame (bodyfixed_to_inertial)
Capability
The body-pnt kind: an orbiter or a surface lander around a body chosen by name navigates with one-way pseudoranges from a Walker navigation constellation around the body (two-body orbits with the body's J2 secular drift, an unknown user clock) and an optional clock-free two-way range from the Earth, the Earth's direction and distance from the analytic ephemeris at every epoch; lines of sight blocked by the body's sphere and a surface elevation mask; per epoch the constellation-only GDOP and PDOP (orbit::dop), the formal position uncertainty with and without the Earth row, and seeded Gauss-Newton fixes (batch_ls::gauss_newton) each way, with the body-to-Earth light time and round trip
Oracle
Closed forms and self-consistency: Kepler's third law for the relay period, the orbit radius and the surface radius held over the run, adding a measurement never increasing the formal uncertainty, and the seeded errors agreeing with the covariance. The relay orbits ignore third bodies (Jupiter's pull on relays around Europa is not modelled), the noise is Gaussian at stated levels and the measurement model is instantaneous, so this is not validated against any mission's navigation data
ValidatedWalker constellation geometry and the published nominal slots of GPS, Galileo and GLONASSconstellation (WalkerSpec::elements,gps_slots,galileo_walker,glonass_slots,GPS_BASELINE_SLOTS,GPS_EXPANDABLE_SLOTS)
Capability
The constellation-design kind's generators: Walker delta and star patterns in the T/P/F convention (node spacing 360/P or 180/P, in-plane spacing 360·P/T, inter-plane phase offset 360·F/T), and the GPS baseline and expandable 24-slot, Galileo and GLONASS presets built from their published nominal elements, placed in an Earth-fixed frame by the Greenwich hour angle their documents state
Oracle
Published agency documents: Galileo Open Service Service Definition Document issue 1.1 (European GNSS Service Centre) Tables 1 and 23, the reference constellation at 2016-11-21 00:00 UTC; GLONASS Interface Control Document edition 5.1 (2008) section 5.2, the slot formula for node longitude and argument of latitude; GPS Standard Positioning Service Performance Standard 5th edition (April 2020) Tables 3.2-1, 3.2-2 and 3.2-3, including the groundtrack equatorial crossing column, which is an independent statement of the Earth-fixed geometry the preset must reproduce. Validates the generators and the transcription; the BeiDou medium-orbit phase and inclined-geosynchronous nodes are not published and are not covered
Tests
constellation::tests::walker_24_3_1_reproduces_galileo_os_sdd_table_23 (all 24 RAAN and mean-anomaly rows to 1e-9 deg); constellation::tests::glonass_icd_slot_formula_is_a_walker_24_3_1 (the ICD slot formula and a…
ValidatedGlobal dilution of precision of the GPS baseline constellationconstellation (coverage,dop_at,NormalAccum)
Capability
The constellation-design coverage engine on the GPS baseline 24-slot preset under the GPS Standard Positioning Service Performance Standard (SPS PS) Appendix B conditions (one sidereal day, 287 five-minute steps, a 4 x 4 deg global grid weighted by the cosine of latitude, all in view, 5 deg mask, one receiver clock): the global HDOP distribution, the PDOP median and the PDOP-at-most-6 availability, globally and at the worst site
Oracle
GPS SPS PS 5th edition (April 2020) Appendix B sections B.3.2.2 and B.3.2.3, the published global-average HDOP distribution of the fully occupied baseline 24-slot constellation (median 0.94, 90 % 1.16, 95 % 1.25, 98 % 1.37, mean 0.96), Table 3.8-1 (PDOP availability) and Table B.3-1 (ensemble PDOP of a degraded constellation, used only as an upper bound), plus a hand-derived closed form for one epoch. The VDOP and PDOP distributions of the full constellation are published only as a figure, so PDOP is pinned one-sided; the worst time-space point (HDOP 2.40 and VDOP 5.22 here, 2.49 and 5.43 published) depends on the grid and is reported, not pinned
Tests
constellation::tests::gps_baseline_global_dop_matches_the_sps_performance_standard (HDOP median 0.940, 90 % 1.165, 95 % 1.255, 98 % 1.370, mean 0.965 against 0.94, 1.16, 1.25, 1.37 and 0.96, bar 0.03 on each fixed…
ModelledCoverage and dilution-of-precision maps for arbitrary multi-constellation designs at scale, around any central bodyconstellation (coverage,ConstellationDesignScenario,beidou_slots,body_by_name)
Capability
The constellation-design kind beyond the two pinned cases: explicit and multi-shell designs, several constellations per run with one receiver clock per constellation, the Earth, the Moon and Mars from the body constants, the BeiDou preset (medium-orbit phase and 118 deg E inclined-geosynchronous crossing modelled), per-cell satellites in view, GDOP, PDOP, HDOP, VDOP and availability, downsampled ground tracks, and a sub-satellite-latitude visibility prefilter that runs a 5 000-satellite design on a 10 deg grid in a fraction of a second
Oracle
Internal consistency: the brute-force elevation scan, the hand-computed and orbit::dop reductions, and the closed-form Walker and geosynchronous geometry. Two-body orbits with an optional secular J2, a spherical body with the local vertical along the radius, geometric visibility only (no signal power, satellite health, terrain or third-body perturbation), and the relative phase between systems taken from different reference epochs, so it is not a snapshot of any date. 0.30 review, a finding (stays MODELLED): the presets meet the published one-sided DOP-availability floors (BeiDou OS Performance Standard 3.0, Galileo OS SDD) at 100 %, but with the engine's PDOP doubled in a scratch copy they still meet them (worst point 99.36 %), so the floors cannot detect even a 100 % error in the quantity this row computes; tightening them after the run would be a post-hoc tolerance. Promotion needs a two-sided comparison of DOP values on identical geometry with a pre-set tolerance. 0.30 round 2 (tests/coverage_dop_full_oracle.rs, pre-registered a4d6f086), a finding (stays MODELLED): Orekit 12.2 positions, elevation, azimuth and ground tracks with gnss_lib_py 1.0.4 GDOP/PDOP/HDOP/VDOP (common clock) and a numpy inverse (one clock per constellation) agree on every bar on the Galileo, GPS, BeiDou, Galileo+GPS (one clock each) and a 1584-satellite Earth design (DOP within 2e-14 relative, identical fix, availability and visibility counts). On the Moon and Mars designs the elements, positions, visibility and tracks agree, but DOP exceeds the pre-set 1e-9 relative bar in cells whose largest PDOP is 1434 or more (up to 8e-6), and 4 cells differ by one fix epoch on numerically singular geometry (engine pivot 1e-10 against numpy rank), so the row is not validated. 0.30 revisions: the TDOP/GDOP reference clock is now the lowest-numbered constellation in view (constellation-multi-gnss-coverage global GDOP max 2.2236 -> 2.1205, mean 1.122 -> 1.131), and per-constellation visible counts are summed as integers
Tests
constellation::tests (prefilter_matches_brute_force_exactly — every map and counter equals a scan with the direct elevation test; five_thousand_satellites_on_a_coarse_grid — 5 000 satellites, prints the run time and…
ModelledA chained mission across scenario kinds on one shared timelinecampaign (run_campaign_detailed,run_chain,presets,extract,to_svg)
Capability
The campaign kind's phases: each phase runs one or more scenarios of existing kinds through run_toml, reads their outputs into named channels (clock time error and guard, mean effective carrier-to-noise density ratio and tracking floor, vertical protection level and alert limit, position error, satellites tracking, alarm flags) by per-kind presets or explicit result paths, places them at the phase start and holds them onto a common grid, with phase boundaries and events; state is handed on by carry (a channel continues from the previous phase's end value), handoff (a previous phase's number written into the next scenario) and end_at (a phase ends at a time a run computed, such as a spoofing monitor's detection time); a campaign hash and a digest over every member result
Oracle
Composition identities against the stand-alone runs of the same kinds: a one-phase campaign reproduces the stand-alone output bit for bit, the carried offset equals the previous run's own last sample, and the phase ended by end_at has exactly the run's detection time as its length. The additive carry across a phase boundary and the zero-order hold are modelling choices, and each phase is as good as the kind that ran it; no chained mission has been checked against a measured one
ModelledParameter sweeps and seeded Monte Carlo ensembles over any scenario kindcampaign (run_sweep,run_monte_carlo,ensemble); sweep (GenericAxis,coords_of,set_dotted_value); inertial (metric_stat)
Capability
The campaign kind's sweep (one to three dotted keys of any kind, on the generic sweep's axis values, optionally an ensemble at every node) and monte_carlo sections: realisation k runs at base_seed + k; each metric reports mean, standard deviation, nearest-rank 5th/50th/95th percentiles and a fixed-seed percentile-bootstrap 95% confidence interval on the mean (inertial::metric_stat); metric units read from the swept kind's own units block
Oracle
The closed form of the phase random walk under white frequency noise, variance q_wf tau (NIST Technical Note 1337), for the statistics machinery, and byte equality with the stand-alone runs for the seeding. It checks the ensemble arithmetic on the engine's own clock model, not an external dataset
Tests
tests/campaign_composition_reference.rs (a_fixed_seed_ensemble_is_byte_stable_and_each_realisation_is_the_standalone_run; the_ensemble_mean_and_spread_match_the_white_fm_closed_form — 200 seeds of a…
ModelledSeveral scenarios under shared conditions, with a combined summarycampaign (run_compose)
Capability
The campaign kind's compose section: named shared values (for example one jammer's power and position) written into each member at the keys it binds them to, every member run, and per metric the smallest, largest and mean value across the members with the member named; metric units must agree across members
Oracle
Composition identity against the stand-alone run. The shared-condition arithmetic is bookkeeping; the physics is that of the member kinds, each with its own row
Tests
tests/campaign_composition_reference.rs (a_one_node_sweep_and_a_one_member_composition_read_the_standalone_number — a one-member composition reads the stand-alone number and result digest)
ModelledBounded optimisation of one metric over campaign parameterscampaign::optimize (run_optimize,nelder_mead)
Capability
The campaign kind's optimize section (CAP-32): one to three dotted scenario keys bounded to [lo, hi] on a lin or log scale, one metric read by result path, goal minimize, maximize or target; method grid (the sweep's axis values) or a deterministic Nelder-Mead simplex in normalised coordinates with box projection and the standard coefficients; a missing or non-finite metric scores worst and stays in the trace; with runs > 1 the metric is the mean over common seeds; every evaluation is a real run of the named kind, listed in trace and counted in the reproducibility digest
Oracle
Internal only: the exact -20 log10(range) closed form of the link equation, an independent brute-force grid over the same kind, and the Rosenbrock function. Round 1's 1-D bar is NOT MET (the pre-registered box excluded the 3 dB solution; the optimiser stopped at the bound), recorded in docs/design/evidence/cap32-optimize/RESULTS.md and not re-scored. The optimum is local to the box and the start; no global-optimum claim
Tests
tests/campaign_optimize_reference.rs (pre-registered 21b64e52 and, after a round-1 protocol miss, 824e9bc1: p2_prime_interior_target_range_matches_the_closed_form — 1-D target solve within 6.1e-15 of r* = r0 10^((m(r0)…
ValidatedBand-limited closed forms for any ranging signal: power in band and early-late code-tracking jitter against published values, with the Gabor bandwidth and offset spectral separation cross-checkednavsignal (sine_integral,bpsk_power_in_band_closed_form,bpsk_gabor_bandwidth_closed_form_hz,dll_jitter_bandlimited_s,dll_jitter_small_spacing_limit_s,bpsk_offset_ssc_closed_form,altboc_15_10_psd,parse_modulation)
Capability
navsignal: the sine integral (sine_integral), the BPSK-R fraction of power inside a double-sided band in closed form (bpsk_power_in_band_closed_form), the band-limited BPSK Gabor bandwidth in closed form (bpsk_gabor_bandwidth_closed_form_hz), the band-limited coherent and non-coherent early-late delay-lock-loop jitter for any unit-area spectrum after Betz & Kolodziejski 2009 (dll_jitter_bandlimited_s) and its vanishing-spacing Gabor bound (dll_jitter_small_spacing_limit_s), the offset BPSK spectral separation coefficient in closed form (bpsk_offset_ssc_closed_form), the Galileo E5 AltBOC(15,10) spectrum (altboc_15_10_psd) and a modulation-label parser (parse_modulation); used by the leo-signal kind for every signal design
Oracle
Externally pinned: 90.3 per cent main-lobe power, the coherent early-late jitter at the stated operating point, the sine-integral table values, and the BPSK self spectral separation coefficient 2/(3 R_c) at zero offset. Internal cross-checks only, not external validation: the band-limited BPSK Gabor-bandwidth closed form and its asymptote, and the offset BPSK spectral separation coefficient at non-zero offsets, both derived here and checked against quadrature. Published closed forms: the BPSK-R sinc-squared spectrum and its 90 per cent main-lobe power (Kaplan & Hegarty, Understanding GPS/GNSS, 3rd ed.); the coherent early-late code-tracking jitter sigma^2 = B_L d / (2 C/N0) chips^2 (Kaplan & Hegarty, section 8) and the band-limited generalisation with its Gabor-bandwidth limit (Betz & Kolodziejski, Generalized Theory of Code Tracking with an Early-Late Discriminator, Part I, IEEE Transactions on Aerospace and Electronic Systems 45(4), 2009); the sine-integral tables of Abramowitz & Stegun (Table 5.1); the spectral separation coefficient as the Fourier transform of the squared autocorrelation (Betz 2001). The AltBOC spectrum is checked for unit area and lobe position only (the Galileo Open Service Signal-in-Space Interface Control Document expression integrates to 8 and is divided by it).
Tests
navsignal::band_limited_tests (sine_integral_matches_tables_and_its_limit — Si(1), Si(pi), Si(2 pi) to 1e-12; bpsk_power_in_band_closed_form_matches_textbook_and_numeric — 0.9028 of the power in the main lobe for…
ValidatedMaximum Doppler of a low Earth orbit navigation satellite, which sizes the acquisition searchleo_signal (max_doppler_hz)
Capability
leo_signal::max_doppler_hz: the largest satellite Doppler on an overhead pass at the elevation mask for a circular orbit, v R_E cos(el) / (R_E + h) times f / c, from each preset's own carrier and orbit altitude; the leo-signal kind adds oscillator and user terms and turns it into Doppler bins
Oracle
Published figures: Xona Pulsar X1 maximum Doppler 32 to 34 kHz (Leclère, Marathe & Reid, ION GNSS+ 2025, https://arxiv.org/abs/2509.19551); Iridium Doppler up to +/-36 kHz (Resilient Navigation and Timing Foundation, Recent PNT Improvements and Test Results Based on Low Earth Orbit Satellites). Earth rotation is left out, which moves the figure by up to about 2.3 kHz; both published figures are met without it.
Tests
leo_signal::tests::max_doppler_matches_published_xona_and_iridium_figures — Xona Pulsar X1 (1593.3225 MHz, 1080 km) 33.2 kHz inside the published 32 to 34 kHz; Iridium (1621 MHz, 780 km) within 0.5 kHz of the published…
ModelledLow Earth orbit positioning, navigation and timing signal designs: code tracking, acquisition, GNSS compatibility and a band trade for any systemleo_signal (LeoSignalScenario,SignalDesign,Component,Shape,code_jitter_m,gabor_bandwidth_hz,detection_probability,threshold_for_pfa,mean_acquisition_time_s,ssc_leo_into_gnss,ssc_gnss_into_leo,jammer_tolerance,gnss_victim,preset,public_signal)
Capability
The leo-signal kind: parameterised signal designs (band, transmit bandwidth, ITU allocation; acquisition, data and pilot components with BPSK(n), BOC(m,n), MBOC or flat spectra, power shares, FDMA sub-carriers, code lengths and data rates) from compiled-in public preset files (Xona Pulsar, Iridium STL, Starlink signal of opportunity, CentiSpace, a generic C-band design and generic UHF/L/S/C/wide-C designs, each with its source URL and unpublished values labelled REPRESENTATIVE) or written inline; per signal the band-limited power spectral density, in-band power fractions, Gabor bandwidth, code jitter against C/N0 and spacing and the ranging accuracy, the acquisition search space, square-law detection probability and mean serial and code-parallel acquisition time, the spectral separation coefficient into and from GPS L1 C/A, Galileo E1, GPS L5, Galileo E5a, E5b and AltBOC with the C/N0 degradation, and the CW, wideband and matched J/S tolerance from the spectrum kind's SSC chain; a band trade of ionospheric delay, free-space loss, ranging accuracy at equal C/N0 and equal EIRP and jammer tolerance; and qualitative shape checks of a design against a described measurement
Oracle
The validated closed forms of the two rows above, the square-law false-alarm identity P_d(snr = 0) = P_fa and Holmes's mean acquisition time, and reduction to the spectrum kind's own SSC chain (the same Jammer::ssc code). The designs themselves are inputs: presets marked REPRESENTATIVE or WORKSHOP are not published specifications, enhanced Feher QPSK, code shift keying and OFDM spectra are approximated, and the shape checks compare shapes, not calibrated levels.
ModelledMulti-band spectrum waterfall (UHF, L, S, C) with designed signals and per-band jammersspectrum (Band::unit_psd,Band::tracked_power_dbw,BandDesign,BandCfg,PanelCfg,SpectrumScenario,Waveform,multi_band_svg); leo_signal (SignalDesign::to_spectrum_band)
Capability
The spectrum kind extended beyond the L band: bands given as a preset signal design (drawn with every component, band-limited to the transmit bandwidth, C/N0 and J/S referred to the tracked component) or as a custom carrier and modulation; extra waterfall panels over any frequency range on the same timeline and colour scale; a wideband (barrage) jammer beside CW, narrowband, chirp and matched noise; per-band J/S and effective C/N0 from the unchanged spectral separation coefficient chain
Oracle
Reduction to the existing spectrum chain and the jamming kind's anti-jam equation (a single-component band gives the same numbers as before the extension) and the signal spectra of the validated rows. The jammer powers, timeline, front-end bandwidths and the designed signals are inputs, and a designed signal is truncated at its transmit band (no out-of-band emission).
Tests
tests/leo_signal_reference.rs (the multi-band waterfall runs with four panels, the UHF, L5-band, S and C jammers each deny only their own band, and a plain-band run is unchanged by the extension)…
ValidatedRain specific-attenuation coefficients for any band a LEO-PNT link usesleo_link::itu (p838_coefficients,p838_k_alpha,rain_specific_attenuation_db_km)
Capability
ITU-R (International Telecommunication Union, Radiocommunication Sector) P.838-3 equations (1) to (5): k_H, alpha_H, k_V, alpha_V from the Tables 1 to 4 curve fits, the path- and polarisation-dependent (k, alpha) and the specific attenuation k R^alpha; used by the leo-pass kind's rain term
Oracle
ITU-R P.838-3 (03/2005) Table 5, the Recommendation's own tabulated k_H, alpha_H, k_V, alpha_V, committed in tests/fixtures/leo_link/p838_3_table5.csv with the PDF URL and retrieval date. The table is rounded to its printed digits, so the bar is that rounding
Tests
tests/leo_link_reference.rs (p838_coefficients_reproduce_table5_to_its_printed_digits: all 115 rows of Table 5, 1 GHz to 1000 GHz, every coefficient within 0.6 of its last printed digit; worst observed 0.51)
ValidatedLong-term slant-path rain attenuation on an Earth-space linkleo_link::itu (p618_rain_attenuation_db,RainPath)
Capability
ITU-R P.618-14 section 2.2.1.1 steps 2 to 10 (slant and horizontal path below the rain height, horizontal reduction and vertical adjustment factors, A0.01 and its scaling to 0.001 % to 5 % of an average year), with the rain height (P.839) and R0.01 (P.837) as inputs because the engine carries neither digital map; applied per band and epoch by the leo-pass kind
Oracle
ITU-R Study Group 3 validation examples file CG-3M3J-13-ValEx-Rev8.3.0 for P.618-14, as tabulated by the ITU-Rpy validation pages, with each site's P.839-4 rain height from the same examples; committed in tests/fixtures/leo_link/p618_14_rain_attenuation.csv with the URLs and retrieval date. Validates the procedure given h_R and R0.01, not the maps
Tests
tests/leo_link_reference.rs (p618_rain_attenuation_matches_the_itu_validation_examples: 56 cases at five sites, 14.25 and 29 GHz, 0.001 % to 1 %, within 1e-4 dB; observed 5e-6 dB)
ValidatedTropospheric amplitude scintillation on an Earth-space linkleo_link::itu (p618_scintillation_db,wet_refractivity,ScintillationPath)
Capability
ITU-R P.618-14 section 2.4.1 steps 3 to 9 (reference sigma from N_wet, effective path length, antenna averaging factor, fade depth for 0.01 % to 50 % of the time), with N_wet an input or computed from temperature, humidity and pressure by the ITU-R P.453-14 expressions; below 4 GHz the leo-pass kind flags the value as an extrapolation
Oracle
ITU-R Study Group 3 validation examples (CG-3M3J-13-ValEx-Rev8.3.0) for P.618-14 scintillation, with each site's P.453-14 N_wet from the same examples; committed in tests/fixtures/leo_link/p618_14_scintillation.csv with the URLs and retrieval date. The wet-refractivity expression from temperature and humidity is not covered by this row
Tests
tests/leo_link_reference.rs (p618_scintillation_matches_the_itu_validation_examples: 42 cases at seven sites, 14.25 and 20 GHz, within 1e-5 dB; observed 5e-7 dB)
ValidatedBuilding entry loss for an indoor LEO-PNT userleo_link::itu (p2109_building_entry_loss_db,BuildingClass); detection (normal_inv_cdf)
Capability
ITU-R P.2109-2 Annex 1 equations (1) to (10) with Table 1: the loss not exceeded with probability P for traditional and thermally-efficient buildings at the elevation of the path at the facade, 80 MHz to 100 GHz; applied per band and epoch to an indoor user by the leo-pass kind (the UHF indoor case)
Oracle
The ITU-R Study Group 3 Clutter and BEL validation workbook values, as transcribed in the reference MATLAB implementation's validation script (github.com/eeveetza/p2109), committed in tests/fixtures/leo_link/p2109_bel.csv with the URL and retrieval date. The workbook prints three decimals
Tests
tests/leo_link_reference.rs (p2109_building_entry_loss_matches_the_itu_workbook: 568 values, 28 GHz at 0 deg and 2 GHz at 45 deg, both classes, P from 1e-7 to 0.998, within 0.001 dB)
ValidatedFirst-order ionospheric delay per band, the ionosphere-free combination and free-space lossleo_link::iono (group_delay_m,iono_free_coefficients,iono_free_noise_amplification); leo_link (fspl_db); linkbudget (free_space_loss_db)
Capability
Group delay 40.3 STEC/f^2 per band; the ionosphere-free coefficients f1^2/(f1^2 - f2^2) and -f2^2/(f1^2 - f2^2) and the noise amplification sqrt(a1^2 s1^2 + a2^2 s2^2) for any band pair; the free-space loss 20 log10(4 pi R f/c) of linkbudget::free_space_loss_db in its kilometre-megahertz form
Oracle
IS-GPS-200 section 20.3.3.3.3.2, which states the L1/L2 group-delay ratio gamma = (77/60)^2 that follows from first-order 1/f^2 scaling, and the Friis transmission formula (Proc. IRE, 1946). Validates the scaling laws, not a slant TEC
Tests
tests/leo_link_reference.rs (first_order_iono_reproduces_the_is_gps_200_group_delay_ratio: gamma = (77/60)^2 to 1e-12, L1/L2 and L1/L5 amplification 2.978 and 2.588…
ValidatedMaximum Doppler a static user sees from a LEO or MEO orbitleo_link::geometry (max_static_user_range_rate,DopplerEnvelope,doppler_hz)
Capability
leo_link::geometry::max_static_user_range_rate: a search over every orbit position and every user on the satellite's horizon of a circular orbit, with the Earth turning, for the largest range rate, plus the orbital and largest Earth-fixed satellite speeds; reported per band as the Doppler envelope by the leo-pass kind and used as the cold-start search window of its low-energy model
Oracle
Leclere, Marathe and Reid, Insights into Xona Pulsar LEO PNT: Constellation, Signals, and Receiver Design, ION GNSS+ 2025, arXiv 2509.19551, Table 1 (static receiver on a 6371 km sphere, Earth rotation the only other effect): 37751.7 / 33628.5 / 31813.4 / 4018.4 Hz on L1. Covers circular two-body orbits
Tests
tests/leo_link_reference.rs (doppler_envelope_reproduces_the_pulsar_paper_table_1: Pulsar IOV 520 km 97 deg, Pulsar FOC polar 1080 km 97 deg, Pulsar FOC inclined 1080 km 53 deg and GPS 20180 km 55 deg; orbital speed…
ModelledA LEO-PNT pass and its per-band link budget against the MEO GNSS satellites in viewleo_pass (LeoPassScenario); leo_link::geometry (design_pass,link_geometry,numerical_check,SatMotion,UserMotion); leo_link::antenna; leo_link::itu (p676_gaseous_attenuation_db); leo_link (system_noise_temperature_k,cn0_dbhz)
Capability
The leo-pass kind: satellites from a designed pass, elements, a TLE through SGP4 or a Walker constellation from constellation-design; per band and epoch the look angles, range, closed-form range rate and range acceleration (checked each run against central differences and reported), free-space loss, EIRP with an isoflux, Gaussian or flat pattern, a patch, hemispherical or isotropic user antenna, ITU-R P.676-10 Annex 2 gaseous attenuation, P.618 rain and scintillation, P.2109 building entry loss, polarisation mismatch, system noise temperature, C/N0, Doppler and Doppler rate, and the first-order delay from a Klobuchar or vertical-TEC slant TEC scaled by the Chapman fraction below the satellite; ionosphere-free pairs with code noise at the pass peak; Galileo or GPS carriers from their interface-document received powers through the same receiver
Oracle
Closed forms and self-consistency: the range-rate and range-acceleration identities against numerical derivatives, the isoflux gain cancelling the range growth, the textbook polarisation-loss limits, the qualitative LEO-versus-GNSS observation (a bell-shaped pass of minutes peaking several dB above flat 44-51 dB-Hz GNSS carriers). The EIRPs and patterns are published received powers turned into an EIRP or representative choices, the gaseous term is the superseded P.676-10 simplified method, and nothing is compared with a measured LEO C/N0
Tests
leo_pass::tests (the_leo_pass_is_a_bell_above_flat_gnss; closed_form_doppler_agrees_with_the_numerical_derivative: under 0.01 Hz and 0.01 Hz/s; indoor_uhf_suffers_less_building_loss_than_c_band…
ModelledNamed LEO-PNT system presets with stated sources, and a system-agnostic engineleo_link::presets (SystemPreset,BandPreset,all,by_id,gnss_meo); leo_pass (resolve_bands)
Capability
leo_link::presets: generic multi-band (UHF, L, S, C), generic C band, Xona Pulsar X1/X5, Iridium STL, Starlink as a Doppler-only signal of opportunity, CentiSpace and an optional Celeste IOD preset, each with its source marked PUBLIC (with URL), REPRESENTATIVE or WORKSHOP, the workshop one in the optional Celeste IOD preset file, compiled in only when that file exists; every band and orbit overridable in a scenario, and a scenario may name no preset at all
Oracle
Transcription of the cited public figures and a consistency check: an EIRP set from a published minimum received power predicts the published maximum within 0.5 dB with a flat pattern. That is a check of the range spread, not a validation of any satellite's antenna; representative presets make no claim about a real system
ModelledLow-energy positioning: time to first fix and energy per fix against duty cycleleo_link::energy (fix_budget,duty_curve,IotReceiver,IotSignal); leo_pass (iot_section)
Capability
leo_link::energy and the leo-pass kind's [iot] section: coherent integration capped by the Doppler rate, square-law non-coherent sums to a detection threshold, a Doppler-bin search over the orbit's static-user envelope (cold) or an aided window (hot), parallel or serial code search, navigation-message time on a cold start, energy per fix and battery life across fix intervals, for each LEO band and the GNSS signal
Oracle
Internal consistency of a stated design model: every power, threshold and search choice is an input, and the model ignores bit-edge ambiguity, missed detection and re-acquisition. Not compared with a measured receiver
ValidatedGlobal-average signal-in-space range error weights for any orbit altitudeleo_navmsg::sisre (sisre_weights,StatsAcc)
Capability
leo_navmsg::sisre::sisre_weights: the radial weight w_R and the squared transverse weight w_AC^2 as averages of cos^2 and half sin^2 of the line-of-sight nadir angle over users uniformly distributed on the visible Earth cap above an elevation mask, for any orbit radius, by Simpson integration; used by every leo-navmsg SISRE figure (at 510 km, w_R 0.46 and w_AC^2 1/2.5)
Oracle
The published SISRE weighting-factor table of Montenbruck, Steigenberger and Hauschild (2018), Multi-GNSS signal-in-space range error assessment - Methodology and results, Advances in Space Research 61(12):3020-3038, doi 10.1016/j.asr.2018.03.041: GPS 0.98 and 1/49, GLONASS 0.98 and 1/45, Galileo 0.98 and 1/61, BeiDou MEO 0.98 and 1/54, BeiDou IGSO/GEO 0.99 and 1/126. The same average evaluated at LEO radii gives the LEO weights; no published LEO table is pinned, so the LEO values are the validated method applied, not separately validated
Tests
leo_navmsg::sisre::tests (weights_reproduce_the_published_meo_and_geo_table — GPS 0.979 and 1/48.9, GLONASS 1/45.0, Galileo 0.984 and 1/61.1, BeiDou MEO 1/54.2, geostationary 0.992 and 1/126.2 at a 0 deg mask, each…
ValidatedGalileo ICD broadcast-ephemeris user algorithm as the base of a LEO navigation messageleo_navmsg::elements (kepler_point,kepler_frame,ephemeris_at,sat_state)
Capability
leo_navmsg::elements::kepler_point: the Galileo OS SIS ICD Keplerian evaluation (Kepler's equation, second-harmonic corrections, Earth-fixed node) written separately from the engine's RINEX evaluator, returning the position with the argument of latitude, inclination and node the along/cross/radial correction frame is built from, and the Liu et al. 2025 extension terms
Oracle
RTKLIB 2.4.2-p13 eph2pos ECEF positions (an independent implementation of the same ICD algorithm) for real BKG/IGS broadcast records of 2024-09-10, both committed with provenance in tests/fixtures/rinex_sp3_interop (NOTICE, rinex_ecef_reference.txt, brdc_multignss_slice.rnx). This validates the Keplerian user algorithm; the along/cross/radial corrections added on top are covered by their own rows, and the Liu terms by the Liu et al. 2025 22-parameter row
Tests
tests/leo_navmsg_reference.rs (the_galileo_user_algorithm_reproduces_rtklib_to_a_millimetre — four real Galileo broadcast ephemerides E10, E19, E24, E25 at tk = 0, +/-600, +/-1800, +/-3600 s, every axis within 1 mm of…
ValidatedCRC-24Q frame check for the LEO navigation messageleo_navmsg::codec (crc24q,encode,decode)
Capability
leo_navmsg::codec::crc24q: generator polynomial 0x1864CFB, initial value 0, no reflection, no final XOR, the check the Kshana LEO frame carries over every preceding byte
Oracle
The CRC catalogue check value for these parameters (reveng.sourceforge.io CRC catalogue, CRC-24/LTE-A: width 24, poly 0x864CFB, init 0, check 0xCDE703), which are the CRC-24Q parameters of RTCM 10403 and the Galileo OS SIS ICD, and the message-type 1005 example frame printed in RTCM 10403
Tests
leo_navmsg::codec::tests (crc24q_matches_the_catalogue_check_value — 0xCDE703 for the ASCII string 123456789; crc24q_matches_the_rtcm_1005_example_frame — the three check bytes 0x360B98 of the RTCM 10403 message-type…
ValidatedLEO broadcast-ephemeris fitter and signal-in-space range error versus fit interval and update periodleo_navmsg::fit (fit_message,lstsq,polyfit); leo_navmsg::truth (TruthOrbit,TruthClock); leo_navmsg (sequence_stats,LeoNavmsgScenario)
Capability
The leo-navmsg kind's fitter and trade: a truth orbit integrated with zonal J2-J6 or EGM2008 gravity to the chosen degree and drag, and a seeded free or steered clock; a Levenberg-Marquardt fit of the Galileo Keplerian set on non-singular elements with weak priors, then along-track, cross-track and radial correction polynomials by linear least squares; the clock polynomial fitted net of the user's relativistic term; SISRE (orbit-only and with clock) over usage periods centred in their fit windows, versus fit interval and update period. Validated part: the 16-parameter Keplerian fit and its orbit-only SISRE against fit interval on real precise orbits, and the integrated J2-J6/EGM2008(20, 70)/drag truth orbit against Orekit 12.2 within 4.6 mm over 6 h (bar 2 cm, same density input); the correction polynomials, the clock fit (seeded free or steered clock) and the update-period trade remain modelled
Oracle
Real precise science orbits as truth (TU Graz IfG/ITSG operational reduced-dynamic orbits, free for any use with acknowledgment; SHA-256 values in the fixture NOTICE) fitted by the 16-parameter set over 20 and 30 min arcs, against the statistics Liu et al. 2025 (Remote Sensing 17(16):2894, CC BY 4.0) publish for the same satellites and days: along, cross and radial RMS each within 1.5x (measured 0.804 to 1.087), SISRE within 1.5x (measured 0.716 to 0.809) and SISRE weights within 0.005 of its Table 2 (measured 0.0007); tolerances fixed before the comparison. The per-component gate gives the test its teeth: zeroing the inclination rate leaves every SISRE ratio inside 1.5x but fails the cross-track gate. The uniform 0.72 to 0.81 SISRE ratio is the paper's: its printed SISRE values use Table 2's weights unsquared rather than its own Eq. 9, which Kshana follows (Montenbruck et al. 2018). The integrated truth orbit (zonal J2-J6, EGM2008 to degree 20 and 70, drag) agrees with Orekit 12.2 (CS GROUP, Apache-2.0) and Hipparchus 3.1, propagated from the same initial state in the same frames with the same density input, within 4.55 mm over 6 h (bar 0.02 m fixed before the run, pre-registration 4e603313). Outside the claim: the correction polynomials, the clock fit and the update-period trade, which keep internal checks (a two-body truth recovered by the 16-parameter set to below 1 mm, and an internal numpy cross-check of the exported messages on GRACE-FO with its CLK1B clock, which was written and run by this project and is not an independent oracle)
ModelledMid-pass LEO navigation message update with a continuity check at the switchleo_navmsg (LeoNavmsgScenario::midpass,worst_case_jump,elevation,geodetic_to_ecef)
Capability
The leo-navmsg midpass-update analysis: the highest pass over a user within a search span, messages on a global schedule with usage periods centred in their fit windows, and at each switch the 3D position jump, the clock jump, the user pseudorange jump, the largest jump over any visible line of sight (closed form over the nadir cone), the range error before and after, and PASS/FAIL against a threshold
Oracle
Internal consistency (the jump identity and the cone bound against brute-force sampling). The threshold is an input; no published LEO continuity requirement is pinned
Tests
leo_navmsg::tests::a_mid_pass_update_is_continuous (every switch's range jump equals the change in range error, and a corrected 300 s fit updated every 150 s passes a 5 cm threshold)…
ModelledDocumented binary encoding of the LEO navigation message with a quantisation-error budgetleo_navmsg::codec (encode,decode,write_payload,read_payload,field_table,quantisation_budget)
Capability
leo_navmsg::codec: Kshana's own frame (preamble, version, model, length, payload, CRC-24Q) for the four ephemeris models, the clock polynomial, SVID, issue of data, band, health, a Klobuchar set, NeQuick-G coefficients and UTC parameters, every field with a stated width and step (field_table), quantisation with range refusal, and a budget of the largest position and clock change a half-step change of each field makes over the validity window
Oracle
Round-trip identities and the budget itself. The format is Kshana's own, modelled on the published message components and the Galileo ICD field set; it is not the bit layout of Celeste or any other system (none is public), so there is no external layout to validate against
Tests
leo_navmsg::codec::tests (signed_fields_round_trip_through_the_bit_packer; week8_resolves_near_the_reference); leo_navmsg::tests::every_model_round_trips_through_the_binary_frame (every model within 2 mm of the exact…
ModelledRINEX-4-style and CSV exports of LEO navigation messagesleo_navmsg::text (rinex_export,rinex_import,csv_export,csv_import,calendar,from_calendar)
Capability
leo_navmsg::text: a RINEX-4-style block (> EPH record header, satellite and epoch line with three clock terms, D19.12 broadcast-orbit lines) with system letter L and record types KP16, KRAC, LU22 and APOL, labelled in every header as a Kshana extension not part of RINEX 4.02, and a CSV table with a default column schema and preset schemas, each with an importer
Oracle
Round trips and the calendar arithmetic. RINEX 4.02 defines no LEO navigation records (IGS RINEX 4.02), so the layout is a documented extension with arXiv 2401.17767 cited as prior art; no external reader of it exists to check against
Tests
leo_navmsg::text::tests (calendar_round_trips_and_knows_the_gps_origin; d19_is_rinex_shaped); leo_navmsg::tests::rinex_and_csv_round_trips (every model within 0.1 mm after the text block, the CSV exact)
ValidatedLiu et al. 2025 22-parameter LEO ephemeris model and its fit on real precise orbitsleo_navmsg::elements (kepler_point,Liu22Extra); leo_navmsg::fit (fit_message,fit_kepler,initial_params,formal_sigma); leo_navmsg::truth (TruthOrbit)
Capability
the Liu22 variant of leo_navmsg::elements::EphemerisModel, evaluated by kepler_point with Liu22Extra: the 16-parameter set plus a semi-major-axis rate a_dot, a mean-motion rate n_dot entering as n = sqrt(mu/A0^3) + dn + n_dot tk, transcribed from the paper Eq. 1 (the fit comparison does not depend on this scale convention, so the meaning of a transmitted nDot value is not validated here), and once- and three-per-revolution radius harmonics Crs1, Crc1, Crs3, Crc3 on the uncorrected argument of latitude. Its fit by leo_navmsg::fit::fit_message takes the paper Section 2.3 start and estimator (every correction starts at zero; plain least squares with no priors; the paper 100-iteration cap), but stops on Kshana parameter-convergence rule (every step below 1e-3 of its formal sigma) instead of the paper rule of an RMS change below 0.1 mm. Fitted over 20 and 30 minute arcs of a tabulated precise orbit
Oracle
Measured: TU Graz IfG/ITSG operational reduced-dynamic orbits (free for any use with acknowledgment; SHA-256 values in the fixture NOTICE) of GRACE-A, GRACE-C, Sentinel-2A and Sentinel-6A are the truth, fitted over 20 and 30 min arcs, and compared with the 22-parameter rows (a_dot, n_dot, Crs3, Crc3, Crs1, Crc1) of Tables 4, 5, 6 and 8 of Liu, Su, Xie, Zhou and Qu (2025), Remote Sensing 17(16):2894, doi 10.3390/rs17162894 (CC BY 4.0; numbers cited, not vendored). Bars fixed before the first comparison: along, cross and radial RMS each within 1.5x two-sided, and SISRE within 1.5x, for every satellite and both arc lengths. The round-2 re-run was pre-registered (e4af7a4e) before the engine change and before the held-out orbits were fetched: the paper Eq. 1 and Section 2.3 fit, plus two new held-out days per satellite against the same printed values. Measured: every paper-day ratio 0.755 to 1.138, and every held-out ratio 0.683 to 1.186. Disclosed: the first run 1.71 (Sentinel-2A) and 1.86 (Sentinel-6A) 20 min along-track ratios were known before the re-run. The lowest held-out SISRE ratios sit near the 0.667 lower bound because the paper printed SISRE uses Table 2 weights unsquared (see the 16-parameter fitter row). The pre-registration mis-stated the paper stop rule (the paper stops on an RMS change below 0.1 mm); the pre-registered parameter-convergence rule was kept. Mutation: restoring zero-centred priors gives Sentinel-6A 20 min along-track 1.08 against 0.70 cm (1.54), which fails. Reviewer probe: the pre-0.30 convention (a factor 1/2 on n_dot) leaves every figure unchanged, so the comparison constrains the function space of the 22-parameter model and the quality of its fit, not the n_dot convention. Outside the claim: the ATOMIC ECEF polynomial and the synthetic five-altitude model-comparison table, which have their own MODELLED row
Tests
tests/leo_navmsg_fit_real_orbit_oracle.rs::liu22_model_matches_the_published_sisre (four paper days: along, cross, radial and SISRE ratios 0.755 to 1.138 of Liu et al. 2025, inside 1.5x)…
ModelledATOMIC zero-clock ECEF polynomial and the synthetic five-altitude LEO model comparisonleo_navmsg::elements (EcefPoly,ephemeris_at); leo_navmsg (liu_comparison,LIU2025_TABLE)
Capability
the EcefPoly variant of leo_navmsg::elements::EphemerisModel: a per-axis ECEF polynomial with no clock terms, the relativistic term from -2 r.v/c^2, scored against a steered clock. Also the model-comparison analysis (liu_comparison over LIU2025_TABLE), which fits the 22-parameter model at the paper five altitudes over 20-minute arcs of a Kshana-integrated orbit and prints the result beside the published SISRE
Oracle
ATOMIC model facts from InsideGNSS (6th-order polynomial, about one minute validity, 30 s refresh, clock steered, 24 cm clock error). These are descriptive facts with no published numbers to compare against. The five-altitude table prints Liu et al. 2025 SISRE (8.88, 6.21, 2.87, 2.11 and 0.75 cm at 320, 475, 786, 966 and 1336 km over a 20-minute arc) beside Kshana fit on its own integrated orbit (4.25, 2.87, 2.20, 1.25 and 0.51 cm; ratios 0.46 to 0.77), not pinned, because the paper fitted real precise orbits and the truths differ. The 22-parameter model and fit themselves are validated against real orbits in the row Liu et al. 2025 22-parameter LEO ephemeris model and its fit on real precise orbits. 0.30 revision: the table figures moved with the paper-convention fit (GRACE-A 5.40 -> 4.25 cm, GRACE-C 4.12 -> 2.87, Sentinel-2A 2.95 -> 2.20, HY-2A 2.38 -> 1.25, Sentinel-6A 0.70 -> 0.51)
ValidatedSingle-frequency ionospheric and UTC services of a LEO navigation messageleo_navmsg::services (klobuchar_delay_m,effective_ionisation_level,system_to_utc)
Capability
leo_navmsg::services: the Klobuchar broadcast delay at any carrier (the engine's L1 model scaled by (f_L1/f)^2), the NeQuick-G effective ionisation level Az = ai0 + ai1 mu + ai2 mu^2 with the all-zero and [0, 400] sfu rules, and system time to UTC with the ICD's three leap-second cases
Oracle
GNSSTk 15.3.1 (ARL:UT, LGPL-3.0, commit 55ea3344, run as a separate program through a harness; no GNSSTk code is vendored): getIonoCorr of KlobucharIonoNavData at L1, L2 and L5, getEffIonoLevel of NeQuickIonoNavData for Az with the all-zero and [0, 400] sfu rules, and getOffset of GPSLNavTimeOffset for the UTC drift, within 1e-3 m, 1e-6 sfu and 1e-9 s; and ERFA via pyerfa 2.0.1.5 (BSD-3), which derives UTC from the IERS leap-second table, for the ICD three leap-second cases across the real 2015-06-30 and 2016-12-31 leap seconds including the 23:59:60 reading, within 1e-9 s (GNSSTk does not implement the ICD case (b) day-length rule, so it was not used there, as written in the pre-registration). Inputs are the real IGS BRDC header coefficients. Pre-registered (6b2f6240; amendment c45e0345 committed a 21 602-row subset of the leap-second grid before any Kshana value was compared). The NeQuick electron-density integration is not implemented, and no correction is made for a LEO satellite flying inside the ionosphere
Tests
leo_navmsg::services::tests (klobuchar_scales_with_inverse_frequency_squared — the L1 value is RTKLIB's 6.1278 m; nequick_az_rules; utc_without_an_event_is_tow_minus_leap_seconds_folded_to_a_day…
ValidatedDoppler a ground receiver must handle from a LEO navigation satelliteleo_fusion::geom (EarthOrbit); leo_fusion::doppler (range_rate,doppler_envelope)
Capability
Exact range rate of a satellite whose Earth-fixed position and velocity come from a two-body orbit with the secular J2 drift and the Earth's rotation (leo_fusion::geom::EarthOrbit::state, velocity the exact time derivative of the position), turned into Doppler at a carrier; the largest absolute Doppler, Doppler rate and jerk above an elevation mask over a window (leo_fusion::doppler::doppler_envelope), reported per LEO system by the leo-pvt kind in doppler mode
Oracle
Published figures reproduced from the stated orbit and carrier: Iridium Doppler up to plus or minus 36 kHz (Resilient Navigation and Timing Foundation, Recent PNT improvements and test results based on LEO satellites) and Xona Pulsar X1 maximum Doppler 32 to 34 kHz (Leclère, Marathe and Reid, ION GNSS+ 2025, arXiv:2509.19551). The Iridium figure is a rounded upper bound, hence the 5% bar. The Doppler rate and jerk are reported but not validated: the modelled overhead-pass jerk of Pulsar X1 is about 1.0 Hz/s^2 against the published 1.26
Tests
tests/leo_doppler_reference.rs (iridium_doppler_reaches_the_published_36_khz: one satellite at 780 km and 86.4 deg flown for a day at 1621 MHz over four latitudes, maximum within 5% of 36 kHz…
ModelledPositioning from LEO Doppler, single- and multi-satellite, with clock-drift and velocity statesleo_fusion::doppler (solve,formal_covariance,single_pass_geometry,offset_site,closest_approach); leo_fusion::pvt_kind (run_doppler)
Capability
Batch Gauss-Newton positioning from range rate with analytic partials (d rho_dot / d r = -(I - u u^T)(v_s - v_u)/rho, d/dv = -u, d/d drift = 1), a constant user velocity as an option, a height pseudo-measurement for a surface user, damped steps; the formal covariance at the truth; along-track, cross-track and vertical projections; the single-pass accuracy against the user's cross-track offset; the time of closest approach. The leo-pvt kind's doppler mode adds the error against window length and a Doppler-only signals-of-opportunity mode (the starlink-sop preset)
Oracle
Internal consistency: noise-free recovery, the kinematic identity rho_ddot = (|v|^2 - rho_dot^2 + d.a)/rho differentiated numerically, zero Doppler at closest approach, the mirror solution of a single pass, and seeded errors against the formal covariance. The Starlink figure of Kozhaya, Saroufim and Kassas (NAVIGATION 72(1), 2025: about 2 m in 20 s with three satellites) is a comparison only: the modelled receiver, with a perfect ephemeris and every error in a 30 Hz Doppler sigma, is several times less accurate over the same window, and the paper's receiver and measurement rate are not reproduced
Tests
leo_fusion::doppler::tests (range_acceleration_obeys_the_kinematic_identity_and_vanishes_doppler_at_closest_approach; noise_free_multi_satellite_doppler_recovers_the_user: to 1 mm and the drift to 1e-6 m/s…
ModelledJoint GNSS and LEO pseudorange positioning with inter-system biases and per-signal error modelsleo_fusion::joint_pvt (solve,dop,code_sigma_dll_m,SystemClock); leo_fusion::system (SystemCfg,System); leo_fusion::pvt_kind (run_joint)
Capability
Weighted least-squares Gauss-Newton fixes over any mix of systems, one receiver clock per system (the inter-system bias estimated with the position) or a known broadcast offset on the reference time scale; per-measurement sigmas supplied by the caller, by default the engine's delay-lock-loop thermal noise (navsignal::dll_code_jitter_chips) at the C/N0 of the elevation combined with the signal-in-space range error; unweighted DOP (GDOP, PDOP, HDOP, VDOP, TDOP) in the local geodetic frame, formal east-north-up sigmas and the bias sigmas; the leo-pvt joint mode compares GNSS-only, LEO-only and fused fixes and sweeps the DOP against the number of LEO satellites added
Oracle
A hand-derived four-satellite DOP (one satellite at the zenith and three on the horizon 120 deg apart, whose normal matrix inverts by hand), the structural identity that a system with its own clock and one satellite adds no geometry, noise-free recovery of position, clock and bias, and seeded errors against the covariance. Internal checks only; not compared with a receiver's output. 0.30 external comparison, a finding (stays MODELLED): on a multi-GNSS IGS day (ABMF 2018-05-13, GPS and Galileo C1C, broadcast ephemeris, Klobuchar) the inter-system bias agrees with RTKLIB v2.4.2-p13 rnx2rtkp (median difference -0.11 ns after the 0.30 broadcast group-delay fix, -0.35 ns before; bar 1 ns), but only 93.6 % of 280 epochs (93.2 % before the fix) are within 3 m of the ITRF2020 coordinate against the pre-registered 95 %. RTKLIB itself reaches 88.2 % on the same data, so the bar is beyond single-frequency broadcast positioning at this station and day rather than a solver discrepancy; the comparison test is committed ignored. 0.30 round 2, stricter inputs at the unchanged 3 m / 95 % / 1 ns bar, a finding (stays MODELLED): with the Galileo E1/E5a ionosphere-free pair (pre-registered ca98316f) 79.9 % of epochs are within 3 m (RTKLIB 67.9 %); with precise CODE MGEX orbits, clocks, satellite antenna offsets and code biases (pre-registered 0d7f6e5f) 94.2 % (RTKLIB 94.1 %), and the inter-system bias differs from RTKLIB by -1.03 ns because RTKLIB single mode ignores satellite antenna offsets (+0.04 ns when both omit them)
ModelledPrecise point positioning convergence with GNSS only and with LEO augmentationleo_fusion::ppp (run,PppScenario,LI_2019)
Capability
A float PPP extended Kalman filter on ionosphere-free code and phase: static coordinates, a white receiver clock, one inter-system bias per extra system, a random-walk zenith wet delay mapped by 1/sin(elevation), a float ambiguity per satellite arc added at rise and dropped at set, and the precise orbit-and-clock error common to a satellite's code and phase, processed as one correlated two-row update in a Joseph-stabilised form; measurements simulated from the same model; convergence time (errors below the thresholds until the end of the run), error curves and the run-averaged position NEES for GNSS only and each LEO case, run as the leo-ppp kind
Oracle
The NEES chi-square test is an internal consistency check (truth and filter share the model). The trend is compared with Li et al., LEO constellation-augmented multi-GNSS for rapid PPP convergence, J. Geod. 93:749-764 (2019), doi 10.1007/s00190-018-1195-2: multi-GNSS 9.6 min shortened to 7.0, 3.2, 2.1 and 1.3 min with 60, 96, 192 and 288 LEO satellites. The bundled scenario gives 7.4, 4.8, 3.2, 2.7 and 2.3 min with its own representative 1000 km, 60 deg shells, noise and stations; the paper's constellations, noise and convergence definition are not reproduced, so this is a MODELLED consistency of the trend (monotone shortening, the largest constellation under a third of the GNSS-only time), not a validation
Tests
leo_fusion::ppp::tests (the_filter_is_consistent_over_monte_carlo_seeds: 24 seeds, run-averaged NEES mean in [2.4, 3.6] and at least 75% of epochs inside the two-sided 95% chi-square band; measured with 150 seeds, mean…
Validated5G non-terrestrial-network positioning accuracy from signal bandwidthleo_fusion::ntn (gabor_bandwidth_flat_hz,gabor_bandwidth_bpsk_hz,gabor_bandwidth_numeric_hz,rms_duration_numeric_s,toa_crb_sigma_m,doppler_crb_sigma_hz,frequency_crb_sigma_hz,NtnScenario::geometry,NtnScenario); leo_fusion::joint_pvt (formal_sigma_enu)
Capability
The Cramér-Rao bound on time of arrival from the root-mean-square (Gabor) bandwidth, c/(2 pi beta sqrt(2 (C/N0) T)), for a flat OFDM spectrum (beta = B/sqrt(12)), a band-limited BPSK spectrum (closed-form numerator) or any spectrum by quadrature, and on the frequency of a complex tone, sqrt(3/(2 pi^2 (C/N0) T^3)); downlink time-of-arrival fixes with an unknown receiver clock and a single-satellite Doppler fix over a pass in the 3GPP n256 MSS S band, run as the ntn-positioning kind
Oracle
Published worked values (P1): Bachl, Lei and Nabeel, arXiv:2608.10270 (2026), single-SSB bounds of NR NTN reproduced from the 3GPP TS 38.211 SSB map through Kshana own numeric functions (W_rms 1.9616 vs 1.96 MHz, sigma_t 38.755 vs 39 us, sigma_f 100.79/87.28 vs 101/87 Hz, sigma_rho 0.5970/0.5831 vs 0.60/0.58 m, each within half a unit of the printed last digit); Betz 2001 Table 1 BPSK-R RMS bandwidths (1.128/3.477 vs 1.1/3.5 MHz, supplementary); the time-of-arrival and Doppler fix covariances agree with numpy 2.4.6 numpy.linalg.inv on the exported geometry (P2; 5e-15 and 1.2e-5 relative, bars 1e-5 and 1e-3). Pre-registered (5351f9f0; part C 8220faaa). The engine fix joint_pvt::formal_sigma_enu (the bound evaluated at the true position, not the noisy fix) was written after the first 200 kHz run failed at 6.5e-5 and was re-run at the unchanged tolerance. The flat-spectrum and complex-tone forms the scenario runs (gabor_bandwidth_flat_hz, doppler_crb_sigma_hz) reach the P1 check through in-crate identity unit tests. Bounds on a multipath-free channel, not achieved accuracy. 0.30 revision: ntn-5g-positioning toa_median_sigma_3d_m 2.0903959 -> 2.0903962 m (NR 5 MHz) and 25.938613 -> 25.938839 m (NB 200 kHz)
Tests
leo_fusion::ntn::tests (the_numerical_rms_bandwidth_matches_the_closed_forms; the_range_bound_scales_as_one_over_bandwidth_and_root_cn0: 0.4157 m by hand at 5 MHz, 45 dB-Hz, 0.1 s…
ModelledLEO-assisted time transfer to UTC against C/N0 and the receiver oscillatorleo_fusion::timing (simulate,utc_offset_s,system_to_utc_s); leo_fusion::pvt_kind (run_timing)
Capability
Clock measurements from every LEO satellite in view at a known position, combined by inverse variance from the pseudorange noise model, filtered by a two-state (phase, frequency) Kalman filter whose process noise is the oscillator class's white and random-walk frequency noise (slot_timing::ClockNoise, clock_state::ClockClass; the coast variance of holdover::coast_phase_variance), a seeded truth clock from the same levels, and the IS-GPS-200 section 20.3.3.5.2.4 system-time-to-UTC expression with a broadcast-offset uncertainty added as a per-run bias; the leo-pvt timing mode sweeps oscillators and C/N0 offsets
Oracle
Internal consistency: the published offset expression evaluated by hand, the normalised error over 40 seeds, and the floor set by the offset uncertainty. The NIST figure for Iridium timing receivers with a miniature atomic clock (under 40 ns from UTC(NIST) over 40 days) is a comparison only; the bundled Iridium scenario gives 10 to 11 ns RMS at nominal power with its representative 5 m range error. Flicker frequency noise and ionospheric delay are not simulated
ValidatedLEO coverage and dilution of precision for polar and Arctic users against MEO GNSSleo_fusion::polar (latitude_sweep,latitude_sweep_at,element_sets,satellite_states_at); leo_fusion::joint_pvt (dop); sgp4 (MeanElementSet::from_earth_fixed,SgpOrbit,teme_to_itrs_matrix)
Capability
Satellites in view and median PDOP, HDOP and VDOP with availability against latitude, over sampled longitudes and epochs, for the MEO GNSS systems alone, the LEO systems alone and all of them, each system with its own mask and clock model, run as the leo-pvt polar mode; the orbits are SGP4/SDP4 element sets at the sweep epoch (2026-01-01T00:00:00 UTC in the polar mode) carried to the Earth-fixed frame on the validated Earth-orbit path (IAU 2006/2000A, UT1 = UTC, no polar motion)
Oracle
Orekit 13.1.8 (Apache-2.0) with Hipparchus 4.0.3: the engine's GCRS states rotated to ITRF by Orekit (IERS 2010), WGS-84 TopocentricFrame visibility against each system's mask, org.orekit.gnss.DOPComputer for one clock unknown and NumPy 2.4.6 numpy.linalg.inv (P2) for several; with the path checked against python-sgp4 and ERFA (row 'One Earth-orbit path'). Bars fixed before the run (pre-registered d92aef9b; T2 and T3 are the 0.30 round-2 bars). The TEME -> GCRS step frozen at J2000 turns the test red (4739 failures). Geometry only; no scintillation, terrain or signal power. GNSS VDOP rises toward the pole (1.20 at the equator to 1.51 at 89.9 deg for GPS and Galileo in the bundled scenario) while a near-polar LEO shell converges there. History: round 2 (e6227a77) agreed on tabulated two-body states; the D8 round-1 full claim against Orekit's own TLE propagation (896f3c2e) was a finding, Orekit's SDP4 at e = 0 and its IERS 1996 TEME, pinned
Tests
tests/leo_polar_coverage_on_path_orekit_oracle.rs (configurations A and B, 4040 samples: Earth-fixed positions within 6.8e-5 m, bar 0.2 m; in-view counts identical on 4040/4040; median DOP within 2.3e-11 relative, bar…
ModelledNamed LEO PNT systems as optional data presets, each with its sourceleo_fusion::presets (all,by_id,cn0_range_dbhz,noise_density_dbw_hz); leo_fusion::system (SystemCfg::build)
Capability
System-agnostic inputs: every system is Walker shells, explicit elements or a GNSS preset, a signal (carrier, chip rate, bandwidth, received power or C/N0 range), a SISRE, a Doppler sigma and a clock model; presets fill only what a scenario leaves out, one file each with sources marked PUBLIC (with a URL), WORKSHOP or DERIVED and the unpublished values listed as representative: Xona Pulsar X1/X5, Iridium STL, Starlink signals of opportunity, CentiSpace, a representative C-band system, the ATOMIC zero-clock ephemeris model, and one workshop-derived preset in the optional Celeste IOD preset file, withheld by deleting that file and its scenarios
Oracle
Transcription of the cited public values, checked for structure (every public source has a URL, Walker patterns divide evenly, the C/N0 from received power over kT at 290 K), not against an external oracle. Representative values are named in each preset and are illustrative
ModelledOne LEO-PNT system end to end: signal design, pass link budget, navigation message and fused positioning, each stage's output handed to the nextleo_pnt_chain (LeoPntChainScenario); leo_pass (BandCfg::signal,design_tracking); leo_navmsg (LeoNavmsgScenario::broadcast_sisre); leo_fusion::pvt_kind; leo_fusion::ppp
Capability
The leo-pnt-chain kind: a leo-signal design (public preset or inline) sets a leo-pass band's centre, bandwidth, chip rate and EIRP split; the pass gives the tracked-component C/N0 and band-limited code jitter per epoch; a least-squares C/N0 line in sin(elevation), the leo-navmsg message's SISRE (fitted at the pass satellite's orbit, plus a stated orbit-determination term in root-sum-square) and the design's carrier and chip rate go to every LEO system of a leo-pvt joint fix and, optionally, to the LEO cases of leo-ppp; every hand-off is reported with its value and unit
Oracle
Self-consistency of the hand-offs: each handed-on value equals the upstream stage's own output, and mutating an upstream input moves every downstream figure that depends on it. The stages keep their own labels; the chain adds no physics and no external oracle
Tests
leo_pnt_chain::tests (a_sine_line_fit_recovers_its_line); tests/leo_pnt_chain.rs (every_upstream_change_moves_the_downstream_figures: a stronger EIRP, a longer fit interval and a larger orbit-determination term each…
ModelledSpoofing detection by LEO Doppler and pass-geometry consistencyleo_link::spoof (range_rate_position_gradient,line_of_sight,fll_frequency_jitter_hz,doppler_consistency_window,doppler_consistency); leo_pass (SpooferCfg,spoof_section)
Capability
The leo-pass kind's [spoofer] section: a spoofer counterfeits, self-consistently in range and range rate, the signals seen at a claimed position that jumps and/or is pushed from the true one after an onset (per LEO band and for the MEO GNSS signal); the receiver predicts every range rate from the broadcast orbit and an independent prior position and velocity, and tests the measured range rates with a generalised least-squares chi-square statistic over a window of epochs (clock drift a nuisance or with a prior), on the GNSS channels, the LEO channels and all channels, with frequency-lock-loop thermal noise from C/N0 (Kaplan & Hegarty 2006, section 5.6.2); detection probability from the non-central chi-square law at a stated false-alarm probability, detection declared at a stated missed-detection probability
Oracle
Closed forms and identities, not an external measurement: the range-rate gradient against a central difference, the information-form statistic against the dense covariance form, a pure clock drift giving zero, a prior-explained push giving at most |dx|^2/sigma_p^2, the statistic quadratic in a jump. No detection figure is compared with a measured attack; the spoofer is idealised (no power, angle-of-arrival or correlation-peak signature)
ModelledSpoofing detection by cross-band consistency of a multi-band LEO signalleo_link::spoof (cross_band_step); leo_pass (spoof_section: pairs,cross_band)
Capability
For every LEO satellite and pair of ranging bands the [spoofer] section forms the geometry-free pseudorange combination, removes the receiver's ionospheric model and tests its epoch-to-epoch step against the two bands' code noise and a stated tolerated unmodelled ionospheric rate (two-sided Gaussian, exact detection probability); a spoofer that counterfeits some bands and not others, or every band without the ionosphere, is seen at its onset; one that counterfeits every band and simulates the ionosphere is not, which the result states
Oracle
Identities: the onset step of an ionosphere-free spoofer equals the pair's modelled ionospheric delay difference, the threshold is the two-sided normal quantile times the step's standard deviation plus the rate bound. The receiver's ionospheric model is the engine's, so the authentic residual is zero in the mean by construction; no external oracle
ModelledIonosphere sounding: slant TEC from the dual-band delay of a LEO passleo_pass (IonoFreePair::geometry_free_stec_tecu,geometry_free_stec_sigma_tecu); leo_link::iono::geometry_free_stec_tecu
Capability
For every band pair of a leo-pass satellite the report gives the slant total electron content the geometry-free code combination recovers at the pass peak, (P2 - P1) f1^2 f2^2 / (40.3 (f1^2 - f2^2)), and its 1-sigma from the two bands' thermal code noise, so any multi-band LEO-PNT design can be traded as an ionosphere sounder
Oracle
The first-order dispersion identity, exact by construction against the engine's own first-order delay (whose 1/f^2 scaling is VALIDATED against IS-GPS-200 in its own row); the precision is thermal code noise only, without multipath or inter-frequency biases. 0.30 round 2, measured (pre-registered bd9812bd), a finding (stays MODELLED): the CODE global ionosphere map CODG1330.18I against ABMF C1C/C2W code with the CODE biases gives 33 of 36 arc means within 3 TECU (worst +6.84 TECU on an evening arc; per-epoch RMS 3.95 TECU), against a bar of every arc; not validated
Tests
tests/leo_resilience_verticals.rs (the_geometry_free_tec_is_the_pass_slant_tec: equals the pass's slant TEC at the peak to 1e-9 and its sigma equals the RSS code noise over 40.3 (1/f1^2 - 1/f2^2))…
ValidatedLunar frame datum covariance with the stations estimated, against 50-digit extended precisionlunar_frame_campaign,fim
Capability
The seven-parameter Helmert datum covariance of the lunar-frame-campaign scenario with the Earth stations estimated (station 1 anchored), as lunar_frame_campaign computes it in binary64: the joint station-and-beacon Fisher information accumulated from the delay Jacobian, the station block marginalised by a Schur complement through a Jacobi spectral inverse, H = AᵀSA formed with the Helmert design and inverted spectrally. The problem is ill-conditioned by construction (Helmert condition number 2.1e8 to 2.3e8, joint information condition number 1.6e12 to 3.0e13): a 1e-9 relative change in the Schur correction moves the datum sigmas by up to 28 %. Reported: the seven datum sigmas, the Helmert condition number and the weakest direction. CLAIM, NARROWED: the engine's datum sigmas and condition number lie within the a-priori worst-case error bound of a backward-stable double-precision pipeline on this problem; this does not claim they are as accurate as double precision allows (a factorisation route is about a thousand times more accurate on the same inputs, and the opt-in square-root solver about ten million times), and the weakest direction is validated only where its bar is below a right angle (the stations-fixed control)
Oracle
mpmath 1.3.0 (BSD-3-Clause) at 50 significant digits, self-checked at 80 (agreement 1.1e-38 or better), on the engine's committed Jacobian, weights and Helmert design: mpmath's own LU solve for the Schur complement, its inverse for H and its Jacobi eigen-decomposition (P2, an independent numerical library). Pre-registered 826dd676 before the fixture existed, with bars from a first-order backward-error formula: constant (3n+1)n + m times the unit roundoff (Higham, Accuracy and Stability of Numerical Algorithms, Theorem 10.4, the Tier B form), times each quantity's exact first-order sensitivity to the joint information, the Schur complement and H. Measured: datum sigmas within 4.8e-6 on 2026-03-18 (bars 9.6e-2 to 2.2e-1), 9.7e-6 on 2024-01-01 (bars 2.8e-2 to 6.9e-2) and 2.0e-14 with the stations fixed (bars 1.0e-11 to 2.8e-7); condition number within 4.6e-6, 1.3e-5 and 4.1e-14. On the stations-estimated scenarios the bars are worst-case and four orders of magnitude above the measured error, and the weakest-direction bar there exceeds a right angle, so that one check is not counted; on the stations-fixed control the weakest direction agrees to 6.6e-16 rad against a 9.4e-11 bar, measured by the chord between unit vectors (an earlier acos comparator could not resolve angles below 1.5e-8 rad; the bar is unchanged). Diagnostic: NumPy 2.3.5 Cholesky and QR routes reach 9e-9 and 5e-9 on the 2024-01-01 inputs, its inverse and eigen routes 2.9e-6 and 6.2e-6: the engine is within its bound but about a thousand times less accurate than a factorisation route. Validates the linear algebra on the committed inputs, not the Jacobian (the SPICE leg of the campaign row does that) and not the physical accuracy of the illustrative campaign
Tests
tests/lunar_frame_campaign_mpmath_oracle.rs (lunar_frame_campaign_datum_matches_mpmath_extended_precision: three scenarios — stations estimated on a new campaign date 2026-03-18 (binding), stations fixed on that date…
ValidatedSquare-root information datum solver against 50-digit extended precisionlinalg_sr,lunar_frame_campaign
Capability
An opt-in square-root information solver (lunar-frame-campaign solver = "srif"; default unchanged) that never forms or inverts an information matrix for the datum: Householder QR (orthogonal-triangular decomposition) of the whitened delay Jacobian as a square-root information filter measurement update, the Earth stations marginalised through the trailing triangular block (whose Gram product is the Schur complement), a second QR of R_bb A for the seven-parameter Helmert datum, sigmas as the row norms of the triangular inverse, and the spectrum (condition number, weakest direction) from a one-sided Jacobi singular value decomposition. Inner products are compensated with a Dekker-split error-free product, so every operation is a correctly rounded IEEE-754 basic operation or square root; identical results on every platform are argued from that, not tested across platforms
Oracle
mpmath 1.3.0 (BSD-3-Clause) at 50 significant digits, self-checked at 80 (agreement 1.4e-43 or better), on the engine's committed Jacobian, weights and Helmert design (P2, an independent numerical library: mpmath's own LU solve, inverse, Cholesky and Jacobi eigen-decomposition). Pre-registered 6e7000be before the solver was written, with bars from a first-order backward-error bound in square-root form: Householder QR constant mn plus (3n+1)n times the unit roundoff (Higham, Accuracy and Stability of Numerical Algorithms, Theorem 19.4 and Chapter 8) times exact sensitivities. Measured: datum sigmas within 3.0e-13 on 2026-06-09 (bars 4.1e-7 to 5.3e-7), 1.9e-15 with the stations fixed (bars down to 4.8e-12), 3.8e-13 on 2026-03-18 and 2.3e-13 on 2024-01-01, where the default spectral solver is off by 4.8e-6 and 9.7e-6; condition number within 4.9e-13; weakest direction within 1.4e-13 rad (control 1.7e-16 against 5.1e-10). LIMIT OF THE CHECK: the bars certify only datum errors below about 2e-7 relative; the measured 1e-13 figures are observations, not what the bars certify. The check rejects the default spectral solver (sigmas 9.7e-6 against 2.9e-7 on 2024-01-01; on the binding date 1.3e-6 against 4.9e-7, a margin of 1.3 to 2.6) and a mis-taken triangular inverse (column norms: red), but a flipped Householder sign, an uncompensated dot product, and sigmas from a spectral inverse of the H formed from the square-root factor (3.7e-9) all pass, so it does not show that each square-root step is needed. Validates the opt-in solver's linear algebra on committed inputs, not the Jacobian and not the default solver; bit-reproducibility across platforms is argued from the operations used and tested only as rerun identity on one platform
Tests
tests/lunar_frame_campaign_srif_mpmath_oracle.rs (srif_datum_matches_mpmath_extended_precision: four scenarios — stations estimated on a new campaign date 2026-06-09 (binding), stations fixed on that date (control)…
ValidatedMaritime trust monitors: the arithmetic they rest on (NMEA decoding, geodesic offsets, the kinematic and sensor-residual statistics)receiver_trust (ingest,maritime)
Capability
For a vessel, receiver_trust decodes the NMEA 0183 sentences a receiver and its instruments emit (GGA, RMC, VTG, HDT, VHW, VBW, ZDA, GSV), forms the east-north offset between fixes (maritime::en_offset_m), and from these the statistics its monitors decide on: the implied speed, dead-reckoning residual, implied acceleration and implied turn rate over 30 s windows, the heading-versus-course, speed-log-versus-ground-speed and antenna-altitude residuals (10 s medians), and the spread and rise of the satellite signal-to-noise.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: (1) the decoding of every field the monitors read, against pynmea2 1.19.0 (MIT) on 8 synthetic logs (1,086 epochs), to 1e-9 in the field's unit (latitude and longitude, altitude, geoid separation, HDOP, speeds, course, heading, signal-to-noise; integers and dates exactly; time of day to 1e-6 s), largest differences at the level of the last digit of a double; (2) the east-north offset between two fixes, as distance and bearing, against the WGS84 geodesic of geographiclib 2.1 (MIT) on 223 pairs of fixes with baselines up to 2 km, from -45 to 70 degrees latitude and across the antimeridian, to 0.01 m and 0.001 degrees (the bearing against the geodesic's mean azimuth, the direction of the chord); (3) the monitors' input statistics at 341 scored epochs of a 400 s synthetic voyage with a turn and a position step, against Python recomputations that use geographiclib for distance and azimuth, pynmea2 for decoding and numpy for medians and standard deviations: implied speed to 0.005 m/s, dead-reckoning residual to 0.02 m, acceleration to 5e-4 m/s^2, turn rate to 1e-3 deg/s, the residual medians and the signal-to-noise spread and rise to 1e-6. WHAT IS NOT: the window lengths, allowances and thresholds of the monitors are this project's own rules; the Python reproduces them from the same documentation, so agreement corroborates the implementation of the arithmetic, not the rules. The THS sentence has no pynmea2 class and the ROT sentence is not read; neither is covered. The table from an NMEA talker and satellite number to a RINEX-style satellite id is this project's own (only the number, the talker and the value are compared). The time-step, host-clock and authentication-status inputs have no oracle. All data are synthetic text from receiver_trust::synth; nothing here is a measurement and no statement is made about how any monitor performs on real interference. Bearing: compared with the geodesic mean azimuth. The pre-registered comparison against the forward azimuth at fix 1 failed at 0.001 degrees (77 of 223 pairs, worst 8.3e-3 degrees) because of a definitional mismatch, not arithmetic; the amendment was made after the first run and both results are recorded (the forward-azimuth comparison stays in the test as a descriptive, non-pre-registered bound of 0.01 degrees). Disclosed: the first runs found inputs that exercised too little (acceleration and turn rate were clamped to zero at every epoch, and the antimeridian track never crossed it), a bug in the synthetic generator (no longitude wrap) and, for the bearing, that the registered oracle quantity (the geodesic's forward azimuth at the first fix) differs from the chord direction by about half the meridian convergence (77 of 223 pairs outside 0.001 degrees, worst 8.3e-3); the inputs were corrected and the bearing compared against the mean azimuth at the same tolerance (amendments 1 to 3 in the pre-registration, made after seeing the results and so not blind); the course and heading comparisons were made circular (360 and 0 degrees are one course, amendment 5). Mutation: a 1e-4 scaling of the east offset fails the geodesy comparison.
ModelledMaritime trust score: the aggregation of the monitors' evidence into a 0-100 score and its three bandsreceiver_trust (score,monitors,live)
Capability
Each monitor reduces its evidence to a ratio of statistic to threshold; the score is 100 minus the sum over monitors of weight times a clamped linear ramp of the ratio (0 up to 0.5, 1 at 1.5), clamped to 0 to 100 and rounded to 0.1, with bands at 90 and 55; each epoch lists the monitors that deducted and their points.
Oracle
Implementation against specification, NOT an external oracle: a clean-room Python re-implementation (scripts/maritime_score_cleanroom.py), written by a reader given only docs/RECEIVER-TRUST.md and docs/MARITIME-TRUST.md and not the Rust source, agrees with score_from_ratios on 600 ratio vectors (score, band, points and the order of deductions, difference 0), after the documentation was completed where the first reader found it silent or inconsistent (tie order, rounding, the ratio of a pass-or-fail monitor, two examples whose numbers did not match the formula). That shows the documentation is sufficient and the code does what it says. WHAT IS NOT CHECKED, and why the row stays MODELLED: the rule itself (the weights, the ramp, the band edges, the evidence hold) is this project's own, pre-registered but with no external source, and nothing here shows that a score of 80 or 40 means what a navigator would want it to mean on real interference; no figure of that kind is given. The monitors' thresholds are likewise this project's own (Row A checks only their arithmetic).
Tests
tests/maritime_trust_reference.rs::t5_score_matches_the_clean_room_reading_of_the_documented_rule; receiver_trust::score::tests (monotone in every statistic, adding a monitor never raises the score, healthy statistics…
ValidatedNMEA 0183 training-stream conformance and decoded valuesnmea_synth (track,sentences,scenarios); cli nmea-scenario
Capability
Synthetic bridge NMEA 0183 streams for crew training (kshana nmea-scenario): GGA, RMC, VTG, ZDA, GNS, GSA, GSV, HDT, VBW and the PKSHT marker, generated from a truth track with modelled receiver, sky, jamming and spoofing effects. Text output only; nothing is transmitted.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: all 225,227 sentences of eight synthetic training streams (four library scenarios, seeds 1 and 7; GGA, RMC, VTG, ZDA, GNS, GSA, GSV, HDT, VBW and the PKSHT marker) parse in pynmea2 1.19.0, an independent NMEA parser, with a valid checksum and the expected talker and type. The position, speed over ground, course over ground, heading, UTC time and date, fix status and GGA satellite count it decodes equal the generator's truth track within tolerances fixed before the first comparison (1.5e-6 deg, 0.06 kn, 0.07 km/h, 0.11 deg course, 0.5 deg heading, 0.02 s), run by scripts/gen_nmea_training_ref.py with the fixture in tests/fixtures/nmea_training_ref/. The GGA satellite-count expectation was corrected after the first run to the NMEA convention (00 without a fix); numeric tolerances unchanged from the pre-registration commit. WHAT IS NOT: that any receiver, chart plotter or bridge system accepts the streams; the realism of the receiver, sky, jamming or spoofing models (nominal sky, modelled noise, drag and replay profiles, all MODELLED); HDOP, altitude, and the DOP and GSA/GSV content beyond syntax and counts; the instructor log. Nothing is transmitted, and no anti-jam or anti-spoof performance is claimed.
ValidatedInterference map: accuracy-code meanings, AIS not-available values, grid cell assignment, route length per cell state, and inland maskinginterference_map (adsb::nacp_epu_bound_m,adsb::nic_rc_bound_range_m,adsb::AdsbParams::classify,ais::ais_position_is_valid,ais::ais_speed_kn,grid::Grid::cell_of,route::exposure,land::LandMask::is_inland)
Capability
The reading of the ADS-B navigation accuracy and integrity codes the map uses (the EPU bound a NACp code stands for, the containment-radius bound a NIC code stands for, and the low and good classes the method derives from them); the AIS position and speed not-available values the AIS reader discards; the fixed latitude/longitude grid cell a position falls in; the route length, and its split by cell state (degraded, not degraded, unassessed, not observed), that route exposure reports; and whether a position lies inland of a land polygon by more than a buffer. A cell flagged degraded or anomalous is a statement about reported accuracy fields or implausible positions, not a finding of interference, and nothing here checks that.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: five independent implementations read the same synthetic inputs. (1) pyModeS 2.21 decodes synthetic DF17 operational-status and airborne-position frames; Kshana's NACp EPU bounds agree to 0.5 m (largest difference 0.4 m), its NIC containment bounds for codes 5 to 11 agree to 1 m, and its low and good classes put all 12 NACp codes and every NIC code the oracle produces (0 to 11) on the same side of the thresholds the oracle's bounds imply (0 disagreements). (2) pyais 3.3.0 decodes AIVDM type 1 sentences whose payload bits the script packs from the raw ITU-R M.1371 field values; Kshana refuses exactly the decoded not-available positions (latitude 91, longitude 181) and speed (102.3 kn), and keeps 0 to 102.2 kn as decoded (0 disagreements over 94 frames). (3) shapely 2.2.0 covers() on cell polygons agrees with grid::Grid::cell_of for 1770 points over cell sizes 0.5, 0.25, 1 and 0.07 degrees, including near-edge points, both poles and the antimeridian (0 disagreements). (4) shapely 2.2.0 route/cell intersection with GeographicLib 2.1 WGS84 geodesic lengths agrees with route exposure on 8 synthetic routes, including two that cross the antimeridian: total length to 0.6% (largest difference 0.49%, the sphere against the ellipsoid) and each cell state's length to 0.6% plus 125 m per piece (largest 21% of that tolerance). (5) shapely 2.2.0 contains() with pyproj azimuthal-equidistant distance to the coastline agrees with the 2000 m inland mask on 992 lattice-centre points around four synthetic islands, one with a lake (0 disagreements; 4 points within 20 m of the buffer were excluded and counted). The tolerances were fixed in tests/fixtures/interference_map_ref/PREREGISTRATION.md before any comparison, and not changed after; two defects in the synthetic inputs found by the first run are recorded there. WHAT IS NOT: Kshana does not decode raw Mode S frames or AIVDM sentences, so this validates its reading of already-decoded values, not a decoder. The oracle's NIC containment radii for codes 1, 2 and 4 differ from DO-260B (20, 8 and 2 NM), so only codes 5 to 11 are compared numerically. The (0, 0) position rule, the altitude floor, the source-type filter, the identifier hashing, the thresholds and statistics of the method, and the cell size choice are Kshana's own and are not checked against anything. The route comparison uses the same straight-in-latitude-and-longitude path definition on both sides, so it checks lengths and cell assignment, not that this path is the right one for a vessel or aircraft. Nothing here shows that a flagged cell is interference. Inputs are synthetic; the oracle fixtures are written by scripts/gen_interference_map_ref.py (pyModeS is GPL-3.0 and is used in that offline script only; CI needs no Python).
Tests
tests/interference_map_reference.rs::adsb_accuracy_and_integrity_codes_agree_with_pymodes (12 NACp codes, 59 NIC rows); tests/interference_map_reference.rs::ais_field_rules_agree_with_pyais_decoding (94 frames)…
ModelledInterference source region: a feasible emitter region from aircraft and ship position reportsinterference_map::emitter (estimate,sessions,build_units,profile,GaussianFix::from_tdoa),interference_map::emitter_out (locate),kshana interference-map locate
Capability
A feasible region for where a source of degraded navigation reports can be, from ADS-B (NIC/NACp) or AIS position reports: radio line of sight on a 4/3-earth radius from the report height and an emitter antenna height range, a limited effective range for the unknown power, affected units pulling the posterior toward their reach and unaffected units with line of sight pushing against it, grouped into sessions in time and space. Output is highest-density regions at stated confidence levels, a posterior centroid with a moment ellipse, session start, end, duration, affected area and distinct-aircraft count above the publication minimum, a time-of-day profile and hours per day. Terrain-free; a single source per session; the credibility is conditional on the model. Optional fusion with a TDOA fix from geolocation.
Oracle
None external yet. Synthetic forward-inverse round trips only, from a generator that shares the estimator's model, so they check the code and the calibration of the stated levels under that model and nothing about real emitters. An external check against the published Jammertest 2025 transmitter sites was pre-registered (docs/emitter-region/PREREGISTRATION.md, signed off by the owner before any scored run) and run once (docs/emitter-region/RESULTS.md): ADS-B NOT ASSESSABLE (2 scheduled sessions, fewer than the registered minimum of 4) and AIS NOT ASSESSABLE (no published session), so no bar was tested and the row stays MODELLED
Tests
interference_map::emitter unit tests (projection round trip, horizon figure, synthetic forward and inverse round trips: the known emitter lies inside the stated region at about the stated rate, edge cases, session…
ValidatedTrust telemetry wire formats: Prometheus text exposition and OpenTelemetry OTLP/HTTP JSONtelemetry (prometheus: Registry,render; otlp: build_payload,Exporter (feature otlp); sample)
Capability
GNSS trust (score, band, reasons, gate and, on request, position) as Prometheus text exposition served on /metrics and as OTLP/HTTP JSON metrics, with a cap on reason label cardinality.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: our Prometheus text exposition for an 84-line synthetic stream (label values with quotes, backslashes, line feeds and non-ASCII text, 64 reason series plus an 'other' fold, gate and position gauges) is accepted by prometheus_client 0.26.0's text parser, and the metric families, types, help text, label sets and values it reads equal the registry's state exactly (values by f64 equality). Our OTLP/HTTP JSON for the same registry parses with protobuf 7.36.2 json_format.Parse(ignore_unknown_fields=False) into opentelemetry-proto 1.45.1's ExportMetricsServiceRequest, and every metric name, description, unit, kind, monotonicity, aggregation temporality, attribute, value, start time and time equals the registry's state exactly. Both parsers were shown to reject an unknown field, a mistyped value and a malformed line (scripts/gen_telemetry_formats_ref.py). Schema 1.3 (0.36), pre-registered (e1e65570) before any tool read it: for a 64-line synthetic schema-1.3 stream (geometry statuses, DOPs, labels and signal loss present, null, absent and unknown; deduction classes; unparsable lines) the same two parsers read the geometry series (Prometheus kshana_trust_geometry*, OTLP kshana.geometry.*) and the class label on kshana_trust_reason_active exactly as the registry implies (24 families and 54 samples; 12 metrics and 31 points), promtool 3.5.0 check metrics exits 0 with no lint problem, and the OpenTelemetry Collector contrib 0.135.0 accepts the OTLP/HTTP JSON (status 200, nothing rejected) and its file exporter records the request unchanged (xval/telemetry-1.3, data-gated: the release tarballs are fetched with pinned SHA-256 and their outputs committed). WHAT IS NOT: the HTTP endpoint, delivery, timeouts and reconnection are not checked by an oracle, nor the Collector's prometheus receiver; the registry is read through accessors, so a fault in the registry state itself would not be caught; the OTLP 'committed output is current' check runs only with the otlp feature; the metric names and their usefulness to an operations centre are not validated; CEF and LEEF are not covered (see the next row).
ValidatedTrust telemetry SIEM events: Common Event Format (CEF) over syslogtelemetry (syslog: cef,rfc5424,SyslogSink; time)
Capability
Band-change (or per-epoch) events as Common Event Format payloads in an RFC 5424 envelope over UDP or TCP, with a documented field mapping and CEF escaping.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: against two independent open-source parsers (oracles: pycef 1.11 (PyPI, MIT); the Logstash CEF codec (logstash-codec-cef 6.2.7 in Logstash 8.15.3)), on a corpus of CEF lines this crate wrote (157 before amendment 4 below, 170 after it, 176 after amendment 5): every band change and every epoch of a committed 84-line synthetic trust stream, plus adversarial text in hosts, reasons, time labels and gate names containing equals signs, backslashes, backslash-equals, pipes, line feeds, carriage returns, tabs, runs of spaces, multi-byte and astral characters and key lookalikes, extreme scores and offsets, every band with and without a previous band. pycef parses the lines whose header needs no escaping (153 before amendment 4, 166 after it, 172 after amendment 5) and recovers, exactly, the header fields, the severity, the exact extension key set and every value (numbers by f64 equality), after the CEF specification's unescape rules, which pycef does not apply and which scripts/gen_cef_ref.py applies. The Logstash codec, which unescapes by itself, parses every line (157, then 170, then 176), meets the same bar, and also recovers four header fields that need escaping (a pipe, a backslash and an equals sign in the version string), which pycef's header pattern cannot read, so header escaping rests on the Logstash codec alone. The comparison bar was registered in commits f8e7465a (pycef) and 6b53e33a (Logstash) before either parser was run, and both met it at the first comparison. Five amendments are disclosed. Amendments (1) to (3) touched no comparison rule and no expected value: (1) a corpus-size assertion was corrected from 5 to 4 band changes, a miscount, since the stream has 4; (2) the Logstash runner's pipeline delimiter wiring was fixed after a first attempt that produced no events; (3) the trust score and epoch offset were moved from the CEF integer fields cn1 and cn2 to the floating-point fields cfp1 and cfp2, a wire-format change made in this release, after which the corpus was regenerated and both oracles were re-run with the bar unchanged and met it again. Amendment (4), 0.36, extended the corpus and the expected key set, registered in ea978bfb before either parser ran on the new lines: schema 1.3 inputs add the satellite geometry in the custom fields cs6 (label geometry: status, then ';' and the deduction classes), cfp3 (geometryHdop) and cfp4 (geometryPdop); 13 geometry cases (the corpus to 170 lines) and the three new keys; both met the bar, otherwise unchanged, at the first comparison (pycef 166 of 170 lines parsed, the 4 tier 2 header cases as before; Logstash 170 of 170). Amendment (5), 0.36, appended 6 geometry-label cases (the corpus to 176 lines), registered in 23e712e2 before either parser ran on them: from inputs that carry the geometry label, cs6 gains a third field (';' and the label) and cn1 (geometryUsedSatellites) the satellites used; the Logstash codec's mapping table was read first (no flexString fields, hence cs6 rather than flexString1); both met the bar at the first comparison (pycef 172 of 176, Logstash 176 of 176), and every earlier reference entry was reproduced unchanged. WHAT IS NOT: neither parser type-checks the cfp values (both keep them as text), so the use of the floating-point fields rests on the CEF dictionary rather than on an oracle; no vendor SIEM was run; the extension key names and their mapping are this crate's choice, and the parsers show only that they are read back as written; the RFC 5424 envelope, delivery, timeouts and reconnection are not covered; pycef is a small parser last released in 2020, and the Logstash codec is the more substantial oracle.
ModelledTrust telemetry SIEM events: Log Event Extended Format (LEEF) over syslogtelemetry (syslog: leef,rfc5424,SyslogSink; time)
Capability
Band-change (or per-epoch) events as Log Event Extended Format 2.0 payloads in an RFC 5424 envelope over UDP or TCP, with a documented field mapping.
Oracle
WHAT IS CHECKED: this crate's own exact-string tests of the documented LEEF field mapping, and the RFC 5424 envelope fields; no maintained open-source LEEF parser exists (a search of the Python package index found none), so the format is checked against its own specification read by the same author, not against an external oracle. WHAT IS NOT: whether the LEEF payloads are accepted by any vendor's parser or SIEM; the escaping rules beyond the documented ones; behaviour of a real syslog collector.
A signed, hash-chained evidence pack of a GNSS trust event: SHA-256 over every file and a chain over the manifest, an Ed25519 signature over the exact manifest bytes, and a verifier that names every failure (a changed, removed, renamed, swapped or added file, a reordered or edited manifest, a manifest re-hashed without the key, a re-signed pack, a wrong pinned key, and a timestamp token that is stripped, replaced, junk or over another document).
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: (1) the Ed25519 functions the pack uses to sign and to verify strictly (evidence::bundle::sign_detached, evidence::verify::verify_detached) reproduce the public keys and signatures of RFC 8032 section 7.1 test vectors 1, 2, 3 and 1024 byte for byte, accept them, and reject every flipped bit of vector 1 and every seventh bit of the others and a non-canonical S+L signature; the vectors are parsed from the RFC text and reproduced by two independent implementations (cryptography/OpenSSL and PyNaCl/libsodium) in scripts/gen_ed25519_rfc8032_ref.py. (2) evidence::bundle::sha256_hex, the hash behind every pack hash and the chain, reproduces the FIPS 180-4 SHA-256 examples (abc; the 448-bit message; one million a), the empty message and hashlib digests of a 0 to 130 byte ladder across the padding boundaries, exactly. (3) A verifier written in Python from docs/EVIDENCE-PACKS.md alone (scripts/evidence_verify_cleanroom.py: hashlib and cryptography; its author read no Rust code) agrees with the Rust verifier on 71 synthetic intact and tampered packs, exactly, on the verdict, the set of failures with file names, the signature verdict, and the chain head recomputed from the files and the one the manifest records. First run 64 of 71 agreed; the disagreements were one specification ambiguity (checking a pinned key against an unreadable manifest), one oracle interface deviation and one incompletely built case; the specification sentence was clarified, the oracle re-implemented from the text and the case completed, with no tolerance or expectation changed (recorded in commit 468afc78). WHAT IS NOT: the clean-room verifier is an independent implementation by a different author, not a third-party authority, and it shares the specification's choices, so it corroborates this implementation of the rules, not the rules; the vectors validate the primitives, not the pack format's design. In this row the RFC 3161 token is read for its message imprint only and the test tokens are synthetic and unsigned; the authority's signature and chain are verified, opt-in (kshana evidence verify --tsa-ca), by the separate row 'RFC 3161 timestamp-token signature and chain verification'. The tamper cases are synthetic. Nothing here is a legal opinion on, or an attestation of, any pack.
Tests
tests/evidence_crypto_reference.rs (RFC 8032 section 7.1 vectors 1, 2, 3 and 1024; FIPS 180-4 SHA-256 examples; exact); tests/evidence_pack_reference.rs::kshana_and_the_clean_room_verifier_agree_on_every_case (71…
evidence::tsa::verify_token, and opt-in kshana evidence verify --tsa-ca <anchors.pem> for a pack's timestamp.tsr: an RFC 3161 TimeStampResp or bare TimeStampToken (CMS SignedData over a TSTInfo) checked offline against trusted anchors the way openssl ts -verify checks it: the signed attributes' content type and message digest over the TSTInfo; the signature over the signed attributes (RSA PKCS#1 v1.5, RSASSA-PSS, ECDSA P-256 and P-384; RSA public operations on crypto-bigint); the ESS signing-certificate v1 (SHA-1) or v2 hash naming the signer certificate, found among the token's and the caller's untrusted certificates; the signer's critical, sole timeStamping extended key usage; the chain to a trusted self-issued anchor with each certificate's signature, CA status and key usage, and every certificate valid at the token's genTime; and the message imprint against the data. Not checked: revocation (CRL, OCSP), name and policy constraints, the nonce
Oracle
OpenSSL 3.0.13 (openssl ts -verify -attime genTime, and openssl cms -verify -cades -purpose timestampsign plus the TSTInfo imprint for RSA-PSS tokens), run by tests/fixtures/evidence_tsa/run_openssl.sh on the 27 cases of cases.tsv: RSA PKCS#1 v1.5 with SHA-256 and SHA-512 and ESS v1 and v2, an ECDSA P-384 root over a P-256 intermediate and P-256 and P-384 TSAs, RSA-PSS, a token without the signer certificate with and without it supplied, CAdES-signed tokens, a flipped signature byte (three algorithms), a changed TSTInfo (two), a flipped intermediate signature, an ESS twin certificate, wrong anchors (four), an expired TSA certificate, a non-critical and a wrong extended key usage, and wrong data (three). Pre-registered in tests/fixtures/evidence_tsa/PREREGISTRATION.md (committed alone, d1886c6d, before the verifier existed; the OpenSSL verdicts re-run afterwards, byte-identical). First run: all 27 verdicts agree (B1) and every genTime agrees to the second (B2). Scope, stated narrowly: self-generated test certificates and tokens, made by the same OpenSSL that judges them; OpenSSL is an independent implementation, not a timestamp authority; no public authority's token is vendored (freetsa.org states no licence). Nothing here attests any real timestamp
Tests
tests/evidence_tsa_openssl.rs::every_case_agrees_with_openssl (the pre-registered comparison); evidence::tsa::tests (civil dates and times, base64, the RSA public operation)…
ModelledReceiver-trust interference thresholds calibrated on the JammerTest 2024 datasetreceiver_trust::calibration (run_dataset); receiver_trust::monitors (MonitorConfig::profile,JAMMERTEST_2024_CALIBRATED); realdata::jammertest; examples/jammertest_trust_calibration.rs
Capability
receiver_trust::calibration::run_dataset: the pre-registered calibration of the C/N0-drop, AGC and jamming-indicator monitors on the public JammerTest 2024 dataset (Zenodo 15911589, GPL-3.0-or-later, fetched at run time, never vendored): per folder a 1 s timeline from mon_rf.csv and rinex.csv, the monitors run once, each post-calibration epoch labelled clean or jamming from the folder's attack_log (shifted from CEST to UTC, 60 s guard), every threshold swept over a fixed grid, the smallest threshold with pooled clean false-alarm probability at most 0.01 selected, and leave-one-folder-out held-out counts; a UBX reader cross-check. The selection ships only as the opt-in profile [monitors] profile = "jammertest-2024-calibrated" (cn0_drop_db 1.5, agc_k_sigma 3.5, jam_ind_threshold 29); the default rule and every output without it are unchanged
Oracle
The dataset's published event logs (attack_log windows), as event truth, under the pre-registration xval/jammertest-trust/PREREGISTRATION.md (committed alone, 0987266, before the harness ran). First and only run (xval/jammertest-trust/RESULTS.md): 14 of 24 folders evaluable (7 with jamming), 3,502 clean and 9,560 jamming epochs. Held-out clean false-alarm probability: cn0_drop_db 0.0231 (81/3502) against a 0.02 bar, FAIL; agc_k_sigma 0.0003, pass; jam_ind_threshold 0.0094, pass; union of the three: false alarms 0.0328, detection 0.807 against 0.05 and 0.50, pass. By the registered rule the row is MODELLED (one bar failed). Narrow: one receiver type (u-blox ZED-F9P), one event, scheduled one-minute slots as truth (detection a lower bound; clean means outside the logged windows); calibrated on the CSV logs, and the UBX reader gives more false alarms on two folders. Maritime trust-score bands are not calibrated
Tests
tests/receiver_trust_jammertest_calibration.rs::the_recorded_calibration_reproduces (ignored, data-gated by KSHANA_JAMMERTEST_DIR: the recorded selection, held-out counts, bar verdicts and verdict)…
ModelledJammertest scoring: a receiver log scored per scheduled test (detection latency, missed alarms, false alarms, time to recover) against the published test planjammertest::plan (parse_catalog,parse_plan,import,import_attack_log,events_toml); jammertest::score (score_run); jammertest::dataset (run_dataset); jammertest (score_log,score_pack); examples/jammertest_score_2024.rs
Capability
jammertest::{plan, score, dataset} (kshana jammertest import|score): the NPRA test catalogue and transmission plan (MIT, fetched with a pinned SHA-256, never vendored) read as scheduled truth with the catalogue ids preserved; a receiver-trust run scored per test (first alarm minus planned start, detected/late/missed against a tolerance, time to the first alarm-free stretch, alarms outside every slot and its guard), with a signed evidence pack that carries the scorecard. The schedule is planned slots, not recorded on-air times
Oracle
The plan's slot counts per test area (76, 77 and 45 slots, of which 61, 63 and 32 transmit) were counted independently with a script over the same JSON; the 2024 dataset's attack_log windows as event truth, under xval/jammertest-score/PREREGISTRATION.md (committed alone, 123dc1e, before the scorer ran). First and only run (xval/jammertest-score/RESULTS.md): the registered minimum-data rule FAILED (4 scorable jamming windows from 4 of 12 UBX folders against a minimum of 10 windows from 2 folders: 5 folders dropped by the baseline rule, 3 UBX logs carry only GPS time of week), so bars B1 to B4 are not assessable; nothing was loosened or re-run. v2 (xval/jammertest-score/PREREGISTRATION-v2.md, committed alone, data handling only: GPS-time-only logs placed from the date, the 12 CSV-only folders, a moving-platform session for moving rovers; same bars and minimum-data rule): 14 of 24 folders kept but 7 scorable jamming windows, minimum-data rule failed again, no bar assessable (RESULTS-v2.md). Scheduled truth: a planned slot is not an on-air record, a miss may be a test that never ran, a latency is from the planned start, and clean means outside the planned slots. Detection by measurement-domain monitors says nothing about a low-power spoofer. One receiver type (u-blox ZED-F9P). The 2025 SimulaMet NMEA dataset is not used (no stated licence, subscription-gated)
Tests
tests/jammertest_plan_roundtrip.rs (data-gated by KSHANA_JAMMERTEST_PLAN_DIR: slot counts per test area of the 2025 plan, catalogue ids, TOML round trip); tests/jammertest_score.rs (a synthetic log with a known drop…
ValidatedTest-bench export (kshana bench-export): Earth-fixed, geodetic and local-velocity columns, attitude columns and NMEA 0183 outputinterop::testbench (write_motion_csv,write_nmea,trajectory_of,export); fusion::pack::truth_trajectory
Capability
The motion CSV's Earth-fixed position, geodetic position, Earth-fixed velocity, heading/pitch/roll and UTC columns, and the NMEA GGA/RMC sentences, written from synthetic scenarios (gnss-ins including automotive-urban-canyon, jamming, gnss-sim). Motion and events only; no signal is written and no simulator or receiver performance is claimed.
Oracle
WHAT IS EXTERNALLY CHECKED, stated narrowly: against pyproj 3.8.0 (PROJ 9.8.1), geographiclib 2.1, pynmea2 1.19.0 and scipy 1.18.1, run by scripts/gen_testbench_ref.py on 4 synthetic scenarios, with tolerances fixed before the first comparison (each the writer's rounding worst case): the Earth-fixed columns equal the WGS 84 conversion of the geodetic columns (worst 1.3e-4 m, bar 3e-4) and the reverse (1.3e-9 deg, bar 2e-9; height 1.3e-4 m, bar 3e-4) on every row; the Earth-fixed velocity rotated to north-east-down by PROJ's topocentric operation equals the engine's true NED velocity (7.5e-5 m/s, bar 2e-4); heading, pitch and roll equal scipy's yaw-pitch-roll of the engine's true quaternion (5.0e-7 deg, bar 1e-6); every NMEA sentence parses in pynmea2 with a valid checksum and its decoded position (1.2e-3 m, bar 5e-3), height, speed, course (5.0e-4 deg, bar 1e-3) and time equal the motion values; the UTC column equals independent date-time arithmetic. The SHA-256 of the exact CSV and NMEA bytes checked is pinned. WHAT IS NOT: the scenario's truth trajectory itself (the driving profile and its integration); the conventions (heading is the body yaw, not the course, and differs from it by up to about 77 degrees on automotive-urban-canyon; GGA quality, satellite count and HDOP are placeholders); whether any laboratory simulator reads the files as intended; any simulator or receiver performance. The oracles check conversions and encodings, not the physics of the scenario. Local-frame handling is checked through the velocity rotation and the NMEA speed and course only (the CSV has no local position column).
Tests
tests/testbench_reference.rs (7 tests: oracle pins, export bytes, worst case over every row, position columns, local velocity vs true NED velocity, Euler angles vs scipy, NMEA vs pynmea2); tests/interop_testbench.rs…
ModelledCompliance mapping to public frameworkscompliance
Capability
compliance (kshana compliance-report): a mapping from Kshana outputs to clauses of public frameworks by number only, filled from the runs given; a result counts for a capability only when it carries the evidence the mapping names, an empty set evidences nothing, and the wording is "supports evidence for"; the report is not a finding that any framework is met and not a rating or approval by anyone.
Oracle
References checked against a script-extracted heading index of the cited public documents read 2026-10-10 (tests/compliance_references.rs). WHAT IS CHECKED: 49 cited clause, section, paragraph or heading references exist in the cited document, and 7 cited URLs resolved on the recorded date (2 via a second route); negative controls fail. WHAT IS NOT: that any paraphrase, mapping or gap is correct; EN 16803 clauses (paid, catalogue text only); later revisions; the index is not an independent implementation.
ModelledMarine integrations: Signal K plugin, OpenCPN gate stream and score panel, reference buildreceiver_trust (live: the gate,its TCP listener,PKSHT); outside the crate: integrations/signalk (JavaScript),integrations/opencpn (JavaScript relay,C++ native plugin),deploy/reference-build (systemd units,container)
Capability
The trust score, band, reasons, gate state and receiver-reported position published under navigation.gnss.kshana in a Signal K server with a warn, alarm and normal notification (integrations/signalk); the gated NMEA stream, with the fix marked invalid while trust is collapsed and a PKSHT sentence per cycle, served to OpenCPN over TCP by the gate itself or by a small relay, and a native OpenCPN score panel that reads the PKSHT sentence (integrations/opencpn); and systemd units and a container that run the gate (deploy/reference-build). Where these show output they carry the advisory statement: Advisory software, not type-approved navigation equipment (IEC 61108, IEC 61162): the operator remains responsible for the navigation of the vessel.
Oracle
WHAT IS CHECKED, stated narrowly: that these integrations work against the real software they are written for. EVIDENCE is committed at integrations/signalk/evidence/README.md (with its JSON files) and integrations/opencpn/evidence/README.md (with its screenshots and logs), each with the script that reproduces it. (1) Signal K: the plugin loaded into the pinned signalk-server 2.33.0 npm package, in a private loopback-only network namespace, fed synthetic data. The server's own REST and WebSocket interfaces showed the published paths, the score metadata, and a notification that went absent, warn, alarm, normal (after the hold) and warn again when the data stopped. With the real kshana run by the plugin on the NMEA the server itself received, the notification went warn then alarm and the reported position was published; the server's own navigation.position was written by its NMEA provider, never by the plugin. Re-run against live JSON schema 1.2. (2) OpenCPN: the plugin built against the OpenCPN plugin API header and wxWidgets 3.2.4, loaded into the Ubuntu 24.04 package of OpenCPN 5.8.4 under a virtual display; OpenCPN loaded it, ran its initialisation and handed it the PKSHT sentences of the gated stream, and the panel turned red on the untrusted band. In one clean run, OpenCPN's own ship position stopped updating, and its fix indicator dropped, when the gate marked the fix invalid. A second run fed directly by the gate's TCP listener showed delivery to the plugin but not a clean before-and-after. (3) The relay tests replay a recorded excerpt of the gated stream through a real TCP socket and check what a consumer receives (the bytes unchanged, valid checksums on every sentence including the rewritten ones, GGA and RMC agreeing, one PKSHT per cycle with the gate field where the fix is invalid), and the end-to-end script checks that a consumer of the gate's listener receives exactly the bytes the same gate writes to standard output. (4) The PKSHT parsers (JavaScript and C++) are run over sentences recorded from kshana output and reject a wrong checksum, version or a score outside 0 to 100. (5) The systemd units are checked with systemd-analyze verify (systemd 255, binaries stubbed) and with guards for a shell in ExecStart, the restart policy and the sandbox keys; the container entrypoint's argument handling is checked with a stub binary. WHAT IS NOT: there is no external oracle for any value these integrations show. The trust values themselves are the rows above, and an integration run shows only that the plumbing carries them; that is why this row is MODELLED and not VALIDATED (docs/VALIDATION.md). No real vessel, receiver, long run, other version of either host, other operating system or desktop was used; all data are synthetic text and nothing here is a measurement. The notification's sound and visual handling by a real Signal K client, OpenCPN's toolbar button, preferences and stale state, and packaging for the OpenCPN plugin manager were not exercised. The systemd sandbox keys were checked for syntax only, not run under a live systemd, and the container image build and its health check have no committed evidence here. No detection or false-alarm figure is claimed. Advisory software, not type-approved navigation equipment (IEC 61108, IEC 61162): the operator remains responsible for the navigation of the vessel.
Tests
tests/receiver_trust_live_cli.rs (the gate and its TCP listener, Rust side); integrations/signalk/test/adapter.test.js, integrations/signalk/test/model.test.js and integrations/signalk/test/plugin.test.js (24 Node…
ModelledBOC(1,1) tracking: DLL scaling and bump-jump ambiguity resolution (opt-in)iq::track::boc (BocKind,BocTracking,BumpJump,boc11_dll_scale); iq::track::channel (the bump-jump taps and Channel::bump_jumps)
Capability
The boc field of iq::track::LoopConfig (opt-in, None by default): the three DLL discriminators rescaled to read the code error in chips with unit slope on the sine BOC(1,1) main peak, and a bump-jump monitor on very-early and very-late correlators at one subcarrier half-period (0.5 chip) that jumps the code NCO towards the main peak when an up/down counter reaches N (8 by default). Without the option every output is unchanged.
Oracle
Internal consistency against injected truth and closed forms, under bars pre-registered at 700d95df (docs/design/evidence/boc-bump-jump/). A BeiDou B1C data signal (PRN 19, 5 MS/s) generated by iq::scene: with the option off a side-peak start stays false-locked at the predicted ±(0.5 + d/4) chip (measured ±0.549); with it on, one jump towards the main peak at 0.08 s and a late-window RMS of 0.0008 chip at 45 dB-Hz and 0.0048 chip at 35 dB-Hz; no jump from the main peak at 35, 40 and 45 dB-Hz over 4 s; the scaled EMLP slope 0.94 to 0.99 on the sampled signal (unscaled 3.6 to 3.8). The bump-jump detector alone, on correlator triples with the exact BOC(1,1) noise covariance, made no false jump in 100 000 updates at 25 to 45 dB-Hz and detected 1000 of 1000 side-peak starts at 35 dB-Hz. NOT MET (finding F1): the pre-registered unit-test grid for the scaling includes offsets outside the discriminators' linear range, where the scaled readings miss the offset by up to 48 %. RE-RUN 0.41 (D8a row 5, 19c6b3518): MET on re-run with the opt-in exact inverse of each discriminator on the ideal BOC(1,1) peak (the linear scaling stays the default; every existing output unchanged): all 27 grid points within 10 % (worst 4.6e-15, by construction, stated in advance); sampled-signal slopes 0.995 to 1.011 (linear 0.938 to 0.994); B1, B2, B3 and B5 hold on registered and fresh seeds. WHAT IS NOT CHECKED: field IQ, band-limited front ends, composite MBOC/QMBOC replicas, and any external receiver.
Tests
tests/iq_boc_bump_jump.rs (b1_b2_b3_side_peak_and_false_jumps_on_the_iq_chain, b4_scaling_on_the_sampled_signal, b5_false_jump_rate_and_detection_at_correlator_level); iq::track::boc unit tests…
iq::channel::scint, the Cornell scintillation model (Rice intensity, second-order Butterworth spectrum, phase arg z) driven at a measured window's S4 and intensity decorrelation time, compared on what it generates beyond those two inputs; realdata::scint50 reads the measured log, detrends it and runs the comparison.
Oracle
Measured data under bars pre-registered at 653a2195 (xval/scint-malindi/): Zenodo record 16941224 (INGV, Malindi, 18 March 2023, GPS PRN 26, 50 Hz intensity and carrier phase, CC BY 4.0, SHA-256 checked). 22 qualifying 60 s windows (N1 met). MET: the intensity distribution (Kolmogorov-Smirnov D 0.0415, bar 0.05) and the phase scintillation tied to S4 (median measured/model sigma-phi 0.779, bar 2/3 to 3/2). NOT MET (finding F1): the model's deep-fade fraction is 2.38 times the measured (bar 0.5 to 2) and its mean fade 3.51 times as long (bar 2/3 to 3/2). Finding F2: the registered processing reads the model's wrapped phase, which a self-check exposed; B3 repeated post hoc with the phase unwrapped gives 0.804. WHAT IS NOT CHECKED: other nights, stations, latitudes or signals; receiver noise in the model. The row stays MODELLED.
Tests
tests/iq_scint_malindi_measured.rs (cornell_model_verdict_is_the_recorded_one, the registered verdict asserted as recorded, finding F1; finding_model_fades_are_deeper_and_longer_than_measured…
ModelledCell-averaging CFAR option in IQ acquisitioniq::acq (CfarConfig,ca_cfar,ca_cfar_cells,ca_cfar_pfa,ca_cfar_alpha)
Capability
iq::acq::ca_cfar and ca_cfar_cells (opt-in): a cell of an acquisition grid is compared with the mean of reference cells of its Doppler row (guard and stride lags, wrapping round the period), at a factor alpha set from the closed-form CA-CFAR false-alarm probability for M non-coherent blocks (ca_cfar_pfa, ca_cfar_alpha). The sample-power detector stays the default and every existing output is unchanged.
Oracle
Closed form, under bars pre-registered at 8af30331 (docs/design/evidence/acq-cfar/). The CA-CFAR false-alarm probability for Gamma(M) cells, against Monte Carlo of the exact model: 18 of 18 cases within 3 binomial sigma. On every cell of iq::acq searches (GPS L1 C/A, 2.5 MS/s, 210 000 cells each) at a design 1 %: 0.977 % and 0.969 % on Gaussian noise and 1.006 % and 0.999 % on 2-bit quantised noise (M = 1 and 4; bar plus or minus 20 %); decisions unchanged when the input is scaled. NOT MET (finding F1): on a noise-free 2-bit signal the detection and its cell are right, but 1.28 % of the off-peak cells of the peak's row (bar 1 %) exceed the threshold, at the code's deterministic correlation sidelobes. RE-RUN 0.41 (D8a row 3, 1059c34e8): MET on re-run with the opt-in row reference set (CfarConfig::with_reference: reference cells every stride-th lag round the whole Doppler row): 0 of 2495 off-peak cells above on the registered signal and 3, 4, 0 and 0 on four fresh signals (bar 1 %), all detected in the right bin; C2/C3 0.969 to 1.033 % on seen and fresh seeds; the 16-cell window stays the default and its control reproduces 0.40. One candidate tried, on development PRNs 2-6 (disclosed). WHAT IS NOT CHECKED: other CFAR variants, two-dimensional windows, band-limited front ends, field data.
ModelledSpectrum model: GLONASS, BeiDou and Galileo E6 bands (opt-in)spectrum (default_band,MORE_BAND_NAMES,GLONASS_CHANNELS)
Capability
spectrum::default_band resolves, by name only (MORE_BAND_NAMES), GLONASS L1OF and L2OF on every FDMA channel K = -7 to 6 (BPSK, 0.511 Mchip/s), BeiDou B1I (BPSK(2)), B1C (MBOC(6,1,1/11)), B2a and B3I (BPSK(10)) and Galileo E6 (BPSK(5)), with the carriers and minimum received powers of their interface control documents; the default band set and every existing output are unchanged.
Oracle
Published interface control documents and closed forms, under bars pre-registered at 502af4a7 (docs/design/evidence/spectrum-bands/): GLONASS ICD Edition 5.1, BDS-SIS-ICD B1I 3.0, B1C 1.0, B2a 1.0 and B3I 1.0, Galileo OS SIS ICD Issue 2.1. MET: every carrier, chip rate and received power as transcribed (28 GLONASS channels), BPSK first nulls at the chip rate, the B1C composite equal to MBOC(6,1,1/11) as its ICD's power table gives, the C/N0 path identity, and byte-identical spectrum scenarios. NOT MET: the unit-area bar, which was unattainable as written (finding F1; the areas equal the closed-form in-band fraction), and the per-bin agreement of the IQ layer's B1I and B1C waveforms with the model envelope (1.26 and 1.56 dB against 0.5 dB, finding F2; GLONASS 0.07 dB), because Gold and Weil codes have uneven line spectra; band-averaged the main lobes agree within 0.25 dB. WHAT IS NOT CHECKED: recorded IQ, other signals, intra-system interference.
ValidatedBeiDou B1I primary ranging codes against an independent generatoriq::signals::beidou (b1i_bits,b1i)
Capability
iq::signals::beidou::b1i_bits and b1i: the BDS-SIS-ICD-B1I-3.0 Gold codes truncated to 2046 chips, PRN 1 to 63, as the IQ layer generates, acquires and tracks them.
Oracle
GNSS-SDR v0.0.19 (tag v0.0.19, commit 1ddd4562), an independent open-source implementation of the same ICD: its beidou_b1i_code_gen_int, compiled standalone, generated all 63 codes. Only its program output is committed: tests/fixtures/iq_signals/beidou_b1i_gnss_sdr.txt, a 64-chip prefix and the SHA-256 of each 2046-chip code. That output is the BeiDou B1I PRN codes the public BDS-SIS-ICD-B1I-3.0 defines; no GNSS-SDR source or binary is in the repository. Bar pre-registered at db9dd696 (xval/beidou-b1i-gnss-sdr/): all 2046 chips of all 63 PRNs identical, no inversion, shift or relabelling. Result: 63 of 63 identical; the fixture was regenerated from a fresh v0.0.19 checkout (source file SHA-256 as recorded) and is byte-identical. WHAT IS NOT CHECKED: the NH20 secondary code, the GEO PRN list, the carrier and chip rate.
ValidatedGPS L1C ranging and overlay codes against the IS-GPS-800J printed check columnsiq::signals::gps (l1c_weil,l1c_ranging_bits,l1c_p_bits,l1c_d_bits,l1co_s1_run,l1c_o_bits); iq::signals::tables
Capability
iq::signals::gps::l1c_p_bits, l1c_d_bits and l1c_o_bits: the L1CP and L1CD Weil ranging codes (computed Legendre sequence, Weil index, expansion sequence inserted before chip p) and the L1CO S1 overlay codes, PRN 1 to 63.
Oracle
IS-GPS-800J (01-AUG-2022, distribution statement A), the check columns it prints beside the generator parameters: Table 3.2-2 initial and final 24 chips of every L1CP and L1CD code, Table 3.2-3 initial and final 11 bits of every L1CO code, Table 6.2-1 the length-10223 Legendre sequence; transcribed by scripts/extract_is_gps_800_tables.py from the PDF (SHA-256 c4aad710...). Bars pre-registered at c6063eb2 (docs/design/evidence/l1c-codes/): 252 of 252 columns, 63 of 63 overlays and 10 223 of 10 223 Legendre bits identical; the extraction --check passes; a negative control fails relabelled, inverted and shifted codes. WHAT IS NOT CHECKED: the generator parameters themselves (they come from the same tables), the chips between the printed columns, the TMBOC subcarrier pattern and the L1CO S2 codes.
ValidatedLunaNet AFS-I and AFS-Q primary codes against the LSIS-AFS V1.0 printed check columnsiq::signals::lunanet (afs_i_bits,afs_q_primary_bits,legendre_bits); iq::signals::lunanet_tables; iq::lunanet_afs::check_spec
Capability
iq::signals::lunanet::afs_i_bits and afs_q_primary_bits (signal names lunanet-afs-i and lunanet-afs-q in kshana iq): the AFS-I Gold code (2047 short-cycled to 2046, G2 delayed per PRN) and the AFS-Q Weil primary code (computed Legendre sequence, Weil index, expansion sequence inserted before chip p), PRN 1 to 210.
Oracle
Against LSIS-AFS V1.0 printed tables (marked TBC in the spec): LunaNet Interoperability Specification V5, Applicable Document 1 Volume A: LunaNet Signal-In-Space Recommended Standard, Augmented Forward Signal (LSIS-AFS) V1.0, 29 January 2025 (NASA; PDF SHA-256 986e0795...), the check columns it prints beside the generator parameters: Tables C-1 to C-5 and D-2 to D-6 first and last 24 chips of every AFS-I and AFS-Q code, Table D-1 the length-10223 Legendre sequence; extracted by scripts/extract_lunanet_afs_tables.py (xval/lunanet-afs/oracle/). Bars pre-registered at e9d5e69 (xval/lunanet-afs/PREREGISTRATION.md) before the generator existed: B1 420 of 420, B2 identical, B3 420 of 420. The document marks its code tables {LSIS-TBC} and says they are for test and integration only, so this validates the construction against that edition's tables, not a later one's. A pre-registered claim about the printed G2 initial state (B1b) was stated with the wrong direction and is NOT MET (0 of 210), pinned; the corrected relation (2047 minus the delay) holds 210 of 210 and is recorded as post hoc. WHAT IS NOT CHECKED: the generator parameters themselves (same tables), the chips between the printed columns, the tertiary code, the message layer and the LDPC coding.
ModelledLunaNet AFS tertiary codes and message layer (SB1 BCH, CRC-24Q, interleaver) against LSIS-AFS V1.0iq::lunanet_afs; iq::signals::lunanet (afs_q_tertiary_bits,afs_q_overlay_bits,afs_q_tiered); surface (lunanet_afs_check_spec,lunanet_afs_check_symbols,lunanet_afs_generate)
Capability
iq::signals::lunanet::afs_q_tertiary_bits, afs_q_tiered; iq::lunanet_afs: synchronisation pattern, SB1 BCH(51,8) encoder and 256-hypothesis decoder, CRC-24Q, the 60 x 98 block interleaver, a symbol-stream checker (kshana lunanet-afs-check) and a test-stream generator with placeholder subframes. The rate-1/2 LDPC coding of subframes 2 to 4 is not implemented (its matrices are in a separate annex).
Oracle
LSIS-AFS V1.0 (as above): the length-1499 Legendre sequence (Table E-1, identical) and the tertiary first and last 24 chips (Tables E-2 to E-6: 419 of 420 columns identical; the document prints the same last 24 chips, A14850, for PRN 147 and PRN 151 and the generator reproduces PRN 151's, so the PRN 147 entry is most likely a misprint; the bar of 420 of 420 is NOT MET and the miss is pinned); the Figure 8 worked example of the SB1 code (one example: 0x045 encodes to 0x229F61DBB84A0); CRC-24Q against the catalogue check value 0xCDE703 for ASCII 123456789; the document's own frame arithmetic. Bars pre-registered at e9d5e69. One miss on record and one worked example for the message layer, so the row stays MODELLED. WHAT IS NOT CHECKED: LDPC and the CRC words inside subframes 2 to 4, message content, and any recorded stream
ModelledStandards packs: scenarios scored against bars stated before the runstandards_pack; compliance (standards-pack facts in compliance-report)
Capability
standards_pack (kshana standards-pack): the DHS Resilient PNT framework test profiles for levels 1 to 4, IMO MSC.1/Circ.1575 accuracy levels and integrity evaluation, and ICAO Doc 9849 Appendix F monitor, assess and mitigate packs: each test runs a bundled scenario kind and compares result fields with bars stated in the pack, and tests that need a receiver log are listed as NEEDS INPUT and not scored.
Oracle
None external for the scenario results: every pack test runs a MODELLED scenario kind. The bars are the pack author's stated profile (the DHS framework states no values; the IMO accuracy limits are Table C-1 of MSC.1/Circ.1575, read in full; the ICAO pack follows ICAO's own RFI Mitigation Plan presentation because Doc 9849 is a paid publication and was not read), committed before any scored run in xval-style PREREGISTRATION.md files under standards/. A MET bar says a modelled scenario met a stated bar, nothing about a product
Tests
tests/standards_packs.rs (pinned_outcomes: the status of every test, including the misses L4-T1 and the four IMO accuracy tests); standards_pack::tests
evidence::roundtrip (kshana evidence round-trip) and the kshana.test-conditions/1 receiver-log export of kshana bench-export: a signed pack that records the scenario hash, the stated conditions' hash, the log hash and the score against the stated events, and refuses conditions that disagree with the scoring events or a log whose hash is not the stated one. No signal is written, replayed or scheduled.
Oracle
None external: the pack links hashes and a MODELLED receiver-log score, checked on synthetic data (the exported NMEA of a bundled scenario treated as the receiver's log). No laboratory run is claimed
ModelledIQ tracking engine against GNSS-SDR 0.0.19 on independent recordingsiq::track (Channel,LoopCore,design); iq::cli::track
Capability
The full IQ-layer tracking engine (iq::track through kshana iq track: acquisition hand-off, correlators, NCOs, second-order 15 Hz Costas PLL, first-order 2 Hz carrier-aided EML-envelope DLL, lock indicator and C/N0 estimators) on GPS L1 C/A at 30 to 45 dB-Hz.
Oracle
GNSS-SDR 0.0.19 (GPL-3.0-or-later, the Ubuntu 24.04 package, run as a separate program; only derived numbers committed) tracking the same numpy recordings (independent of the engine), both measured against the analytic truth. Bars pre-registered at b1f25cce, amended before any test recording at 9496b6b0 (a 10 Hz FLL during pull-in only in both receivers, after the registered calibration failed on a wrong GNSS-SDR hand-off, finding F3). MET at 30, 35, 40 and 45 dB-Hz: carrier jitter within 4.3 %, code jitter within 4.6 % and Doppler mean and spread of GNSS-SDR's. NOT MET: the engine's NWPR C/N0 reads 2.3 to 7.1 dB below GNSS-SDR's SNV (finding F4, the known NWPR limit; its M2M4 is within 0.06 dB of truth); the phase lock indicators differ by definition (F5); at 30 dB-Hz GNSS-SDR slipped twice and the engine not (F6). GNSS-SDR's own hand-off locked to a wrong frequency in 3 of 30 calibration runs. WHAT IS NOT CHECKED: other signals, data-bit wipe-off, re-acquisition, field recordings. The row stays MODELLED.
Tests
tests/iq_track_gnss_sdr_oracle.rs (calibration_passes_for_both_receivers, registered_bars_give_the_recorded_verdict, recorded_numbers_are_pinned), on numbers derived by…
ModelledCRPA antenna gains in the jamming kind against an independent numpy recomputationjamming (CrpaCfg,crpa_weights,run_jamming); crpa (null_steering_weights,mvdr_weights,covariance)
Capability
The jamming kind's optional [crpa] section: a per-channel beam toward each satellite (null-steering with the jammer as the null, or MVDR on sigma^2 I + P_j a_j a_j^H), scaling each link's noise by ||w||^2 and the jammer by |w^H a_j|^2 before the existing anti-jam C/N0 equation.
Oracle
An independent numpy recomputation (xval/crpa-jamming/oracle.py, no shared code: numpy steering vectors, numpy.linalg.solve and inv) of every link of the jamming-demo geometry with a 4-element square array, under both weightings. Bars pre-registered at 716de66a. MET: A1 (all 150 scenario results byte-identical without the section, after a fix: the first build added the crpa units to every jamming run's units table), A2 (one element equals no array to 1e-12 dB), A3 on the array noise gain and effective C/N0 (912 link-weightings within 1e-9 dB, worst 2.1e-14 dB) and on the null-steering jammer null (below -200 dB on both sides), A4 (MVDR within 0.1 dB of null-steering at a JNR of 80 dB). NOT MET: A3 on the MVDR jammer gain, 194 of 456 links over 1e-9 dB on 0.36 portable maths (202 on host maths; docs/revisions/STUDIO15-batch4-rebaseline.md) (worst 1.7e-6 dB) on residuals of -100 to -188 dB, where float64 cancellation limits both implementations (numpy's inv and solve disagree with each other on 169 links, worst 7.9e-7 dB); the effective C/N0 these feed is unaffected. WHAT IS NOT CHECKED: mutual coupling, element mismatch, wideband processing, multiple jammers, any antenna hardware. A recomputation of the same equations, not an external oracle; the row stays MODELLED.
Tests
tests/jamming_crpa.rs (a1_crpa_units_only_with_the_section, a2_one_element_is_no_array, a3_verdict_is_the_recorded_one, a4_mvdr_tends_to_the_null_at_high_jnr), against xval/crpa-jamming/expected.json from…
ModelledTerrain shadowing of the jammer in the jamming kind against an independent numpy recomputation on a real SRTM tilejamming_terrain (TerrainCfg,jammer_terrain,knife_edge_loss_db); jamming (run_jamming_sky)
Capability
The jamming kind's optional [terrain] section: the great-circle terrain profile between jammer and receiver on an SRTM .hgt tile with a k-factor Earth bulge, the dominant knife edge (largest Fresnel-Kirchhoff nu) and the single knife-edge loss of ITU-R P.526-15 section 4.1, subtracted from the jammer's received power.
Oracle
An independent numpy recomputation (xval/jamming-terrain/oracle.py, no shared code: its own .hgt read, bilinear interpolation, WGS 84 round trip, profile, bulge, nu and J(nu)) of ten jammer placements around a receiver at Badwater Basin on the committed real SRTM tile tests/fixtures/terrain/N36W117_sub6.hgt. Bars and placements pre-registered at 3b0ebdd4. All MET: additivity (all 150 scenario results byte-identical without the section); every placement's profile points and obstacle equal and nu and loss within 1.6e-10 (bar 1e-9); the loss never rises as the jammer rises from 2 to 4902 m; the loss lowers every J/S by exactly itself and nothing else; J(nu) within 1e-12 dB of the oracle's. Findings: placements across the Black and Funeral Mountains lose 44 to 48 dB; two valley-floor placements lose 8 and 15 dB from the Earth bulge alone at 2 m antenna heights, the single knife-edge model's grazing case. WHAT IS NOT CHECKED: any measured propagation, multiple-edge diffraction, ground reflection, clutter, the DEM's own error. A recomputation of the same model, not an external oracle; the row stays MODELLED.
ModelledAcquisition hysteresis and the front-end quantiser in the jamming kind against independent recomputationsjamming_rx (AcquisitionCfg,AcqState,FrontEndCfg,front_end); jamming (run_jamming_sky); iq::frontend::quant (optimum_step_sigma)
Capability
The jamming kind's optional [acquisition] section (a satellite is acquired at or above an acquisition threshold after a re-acquisition delay and held while at or above the tracking threshold) and [front_end] section (the AGC reading 10 log10(1 + J/N) and the ADC's Gaussian-input quantiser loss taken off every link, flagged as not valid for non-Gaussian jammers).
Oracle
An independent Python recomputation of every acquisition flag (xval/jamming-acq-agc/hysteresis.py) from the engine's per-epoch effective C/N0, an independent numpy Monte Carlo of the quantiser loss (xval/jamming-acq-agc/quantiser_mc.py, its own optimum-step search, 2^22 samples), and the IQ layer's AGC loop. Bars pre-registered at 2ca31ee1, with a supplementary run set at 2ef7427a. MET: additivity (all 150 scenario results byte-identical without the sections); every flag equal on the supplementary runs (1824 links, 732 of them showing the hysteresis); monotone in the threshold; the quantiser loss within 0.0007 dB of the Monte Carlo at 1 to 4 bits (bar 0.02 dB); the AGC reading within 0.012 dB of the AGC loop (bar 0.05 dB). NOT MET: the registered hysteresis runs (jamming-demo at 10 to 30 dBW) were uninformative, with every link lost, and their H2 and H4 are met only trivially. RE-RUN 0.41 (D8b row 7, 0ea0da3df): MET on re-run on a 12-power ladder fixed from the scenario file before any engine run (-34 to -21 dBW): 2614 of 5472 link-epochs in the informative 25-32 dB-Hz band (bar 10 %); H2 and H4 hold; all 5472 flags agree with hysteresis.py, 2626 showing the hysteresis; no engine change. Disclosed: the ladder rule was written after supplement 1 had shown where the thresholds lie. Finding: with a CW jammer a 1-bit ADC loses about 5 dB more than the Gaussian figure the engine applies, and flags as not valid. WHAT IS NOT CHECKED: the acquisition search itself, AGC dynamics in time, any receiver hardware. Recomputations of the same model, not an external oracle; the row stays MODELLED.
ModelledCapped FLL pull-in below the default design's FLL-to-PLL hand-over floor (opt-in)iq::track (LoopConfig,FllGate,design); iq::track::channel (the FLL gate)
Capability
iq::track's opt-in carrier.fll_pull_in_max_s (the fll_pull_in_max_s field of iq::track::LoopConfig): under fll_assist = pull-in, the FLL hands over to the PLL after a fixed pull-in time whatever the PLI, with a matching return dwell until the PLI first reaches fll_off_pli.
Oracle
Internal consistency against injected truth: a synthetic data-free GPS L1 C/A source (PRN 13, 1500 Hz, 4.1 MHz, 8 s, ten noise seeds) whose carrier phase is known, under bars pre-registered at 8aceabcb (docs/design/evidence/fll-handover-floor/). MET: the control (the uncapped default never hands over in 9 of 10 seeds at 33 dB-Hz); additivity (default hash and the existing tracking tests unchanged); pull-in of a 100 Hz-off hand-off at 45 dB-Hz. NOT MET: with a 1.0 s cap, hand-over and hold in every run (16 of 20 at 33 dB-Hz, 8 of 20 at 31 dB-Hz; the rest cycle between forced hand-overs and the FLL's return) and phase jitter within 25 % of the PLL alone (10.8 against 6.1 degrees at 33 dB-Hz). RE-RUN 0.41 (D8a row 4, 74ba62787): with the opt-in carrier.fll_handover_mean_s (the PLL seeded at the mean NCO Doppler of the pull-in's last 0.5 s), still NOT MET: at 33 dB-Hz hand-over and hold 20 of 20 on seen and fresh seeds with jitter within 1 % of the PLL alone (this half met), but at 31 dB-Hz 15 of 20 seen and 6 of 20 fresh, jitter ratio 1.8 and 2.8. Development seeds 1001-1010 recorded, nothing tuned on them. WHAT IS NOT CHECKED: data bits, other signals and sample rates, field recordings. The default is unchanged; the row stays MODELLED.
ModelledWrite a JPL Horizons state table as a CCSDS OEM, a CCSDS OPM or an SPK kernel, and read it backephem_export
Capability
kshana ephem-export (and ephem_export in Python, WebAssembly and MCP): a JPL Horizons vector table is written as a CCSDS OEM 2.0, an OPM 3.0 or an SPK of type 13 or 9 (frame code 1, ICRF). Values and epochs are copied, not computed, and the engine reads the written file back and reports the comparison
Oracle
The engine's own reader on the files it wrote (same codebase). The C7 reference integration has no bar: Orekit 13.1.9 and Basilisk 2.12.2's SPICE read the files back exactly (Orekit's km to m conversion named as rounding further), and ANISE 0.10.6 evaluates the SPK no further from the written states than SPICE does; GMAT is not done until its CI job is green. A one-off check outside the repository found the SPK byte for byte equal to the kernel SPICE's spkw13 writes from the same states. Carries data only; says nothing about any propagation
Tests
tests/ephem_export_capstone.rs (4321 of 4321 CAPSTONE states read back with the same bits from the OEM and the SPK; the SHA-256 of each file pinned through the CLI); the flight-dynamics workflow…
ModelledOptical-navigation measurement model: apparent direction, pinhole projection with distortion, and partialsopnav
Capability
opnav::measure: the apparent direction of a body centre (converged Newtonian light time and stellar aberration, the SPICE CN+S correction) or of a catalogue star (aberration only), projected through a pinhole camera with Brown-Conrady distortion in the OpenCV projectPoints convention at a given attitude quaternion, with the partials of the pixel with respect to the spacecraft position and velocity
Oracle
SPICE (spiceypy 8.3.0, CSPICE N0067) and OpenCV 5.0.0, under bars pre-registered in docs/design/evidence/c5a-optical-navigation-model/: O1 body directions within 0.0016 mas of SPICE CN+S (bar 1 mas), O2 stars within 3e-8 mas, O3 projection within 1.3e-11 px of OpenCV, O4 end to end within 3.8e-8 px (bar 1e-4 px), O6 additivity: MET. O5 NOT MET: 258 of 432 partial elements differ from central differences at the registered 1 m and 1 mm/s steps by more than 1e-6 relative (the largest elements agree to about 1e-7 to 1e-9; a no-bar diagnostic at 1 km and 1 m/s cuts the worst relative difference from 1.0 to 8e-4, pointing to difference round-off, not proven). RE-RUN 0.41 (C5a v2, PREREGISTRATION-v2.md, 755bb1719): O5 v2 MET against an independent complex-step oracle (xval/c5a-opnav-v2/): the analytic partials agree within 2.3e-11 of each row's largest element on 72 measurements, v1's 36 and 36 fresh (bar 1e-6 of the row's largest element, chosen after v1 and weaker than v1's for small elements; the oracle's own values within 3.6e-8 mas of SPICE and 4.6e-11 px of OpenCV); O1 to O4 on the 36 fresh measurements met (3.7e-3 mas, 2.3e-8 mas, 4.7e-11 px, 8.4e-8 px); observer-relative central differences, reported with no bar, still put 144 of 288 rows above 1e-6 (worst 4.0e-5 of the row's largest). v1's O5 NOT MET stands as recorded. WHAT IS NOT CHECKED: image processing, the body's figure or phase, attitude error, any real camera
Tests
tests/c5a_opnav.rs (36 measurements against the committed SPICE and OpenCV oracle output; the recorded verdicts asserted, the O5 miss included); tests/c5a_opnav_v2.rs (the v2 partials against the complex-step oracle…
ModelledCAPSTONE NRHO propagation with DE440 third bodies and one radiation-pressure parameterc3_force
Capability
c3_force::C3Model: the A2 lunar force model with the Sun and Earth from a DE440 kernel plus Venus and the Jupiter system barycentre as point masses (M1), and optionally cannonball radiation pressure k = Cr A/m with Earth and Moon conical shadows (M2). With no kernel and no radiation pressure it is the A2 model bit for bit
Oracle
JPL Horizons reconstruction of CAPSTONE (-1176), under bars pre-registered in docs/design/evidence/c3-capstone-nrho-force-model/. C3-R MET on A2's window W0 (21.66 km end position, 4.24 km periapsis; bars 25 and 10 km). C3-H NOT MET on the held-out windows E, H1, H2 (end position 37.25, 36.23, 49.65 km; periapsis met). C3-S NOT MET (k fitted on E: H1 12.92 km met, H2 57.13 km). A2 is not met on re-run. RE-RUN 0.41 (C3 v2, PREREGISTRATION-v2.md, cc8c300a0): C3-S v2 MET on the fresh span G (Horizons -1176, 2023-07-01 to 2023-09-28) with k estimated per window from its first 78 samples (M3): end position 0.21, 0.60 and 0.73 km and periapsis radius 11, 4 and 87 m on G1, G2 and G3 (bars 25 km and 10 km); G1 is marked low-information (a 10 m or 1 cm/s change in its start state moves the end position by up to 18.1 km) and its verdict counts as registered. M1 alone also meets both bars on G (5.07 to 13.63 km, periapsis 2.07 to 4.68 km), so G is a less demanding span than F; with its own k, H2 is still missed (57.82 km, no bar). v1's C3-H and C3-S NOT MET stand. Nothing here identifies CAPSTONE's real area-to-mass ratio. The reconstruction is the reference, not truth
Tests
tests/c3_force.rs (kernel off equals A2 bit for bit; radiation term against its closed form and zero in both umbrae; the control on W0, gated); tests/c3_verdicts.rs (the recorded verdicts, misses included)…
ModelledImpulsive steps in the cislunar coast and the C2 step detectormission_thread::impulses
Capability
The tli-and-cruise phase's optional impulses list (t_s from the phase start, dv_mps in GCRF) adds each step to the mean state, shortening a step to land on it; without impulses every output is unchanged. mission_thread::impulses::detect flags hours whose one-hour velocity residual exceeds 0.01 m/s and gives one impulse per event hour at its midpoint
Oracle
JPL Horizons reconstruction of Artemis I (-1023), under bars pre-registered in docs/mission-thread/PREREGISTRATION-C2.md: NOT MET on W1 (17.3 km against 2 km; 0.0177 m/s against 0.01 m/s); W2 not run (the published burn times leave 42.8 h, not 72 h). A diagnostic after the verdict shows four of the five steps in the reconstruction carry position jumps no velocity step within one hour can produce. v2 (docs/mission-thread/PREREGISTRATION-C2-v2.md, 036c3f1be): NOT RUN. By the registered rule a fresh 72 h window W3 after W1 needs at least 120 h between published burns (24 h clear on each side); the longest gap after W1 in NASA's published Artemis I timeline is 105.9 h (return trajectory corrections 4 to 5), so W3 does not exist. No v2 code was written and no Horizons state of any v2 period was fetched (xval/mission-thread-c2-v2/README.md); v1's NOT MET stands. The impulse support itself is checked on synthetic data only
Tests
tests/mission_thread_c2_impulses.rs (synthetic only: start impulse equals a changed start velocity bit for bit; a mid-phase impulse equals stopping, adding and continuing; two known mid-hour steps recovered; malformed…
ModelledSpace RF geolocation bounds for a surface emitter seen from low Earth orbit: one satellite, two- and three-satellite formations, and several passes fusedspace_geoloc
Capability
space_geoloc: the constrained Cramer-Rao bound (the emitter on the WGS 84 ellipsoid) of a static surface emitter from one satellite's frequency of arrival over a pass (S1), from a two- or three-satellite formation's time and frequency differences (F2, F3), and from several passes fused (MP), with the noise per measurement from a free-space link budget; space_geoloc::ccrlb_with_receiver_errors adds receiver position and velocity errors as Gaussian nuisance parameters (Cao et al. 2015), and space_geoloc::solve_f3 is a Gauss-Newton estimator for F3
Oracle
Bars pre-registered in docs/design/evidence/m1-space-rf-geolocation/ (results a41eb6d06). All MET: B1, an independent numpy recomputation (xval/m1-geolocation/oracle.py, no shared code) over 60 cases (S1, F2, F3 x P1-P10 x AIS, a 30 dBm jammer): worst relative difference 3.0e-10 and worst orientation 3.0e-11 rad (bars 1e-6), every pass identical; B2, the published case of Cao et al., Chin. J. Aeronaut. 2015, Fig. 1 (25 plot readings, bar 20 %): worst +5.5 % (at a receiver-location error of 0.75 km), 22 of 25 within 1 %; B3, the F3 estimator against the bound (AIS, P1-P10, 1000 trials each, bar 0.9 to 1.25 x DRMS and 99 % converged): ratios 0.979 to 1.019, 1000 of 1000 converged at every point; B4, the fused information equals the sum of the per-pass information to 6.7e-16 relative and DRMS never rises with another pass; B5 additivity. WHAT IS NOT CHECKED: no flown system and no real RF data; no clock or orbit-knowledge error, multipath, AIS slot collisions seen from orbit, ionosphere at VHF or emitter motion; the link budgets are representative, not measured, and the radar classes isotropic (an upper bound), so the bounds are optimistic
Tests
tests/space_geoloc_m1.rs (B1 against the committed numpy oracle output; B2 against the 25 values transcribed in docs/design/evidence/m1-space-rf-geolocation/reference.csv; B3 and B4 in release; B5 additivity)
ModelledDark-vessel and AIS-integrity decisions from RF fixes against AIS reportsm4_logic; m4_dark_vessel
Capability
m4_logic::decide: from an AIS stream and an RF fix stream (no identities on fixes), one class per identity, NORMAL, DARK (fixes with no matching AIS), GAP, POSITION-OFFSET (AIS position inconsistent with the RF fixes) or IDENTITY-DUPLICATE, and a dark or associated flag per fix; v1 decides by single-fix gates and a vote, v2 (opt-in parameters) by fix sequences for dark tracks and a time-consistent test for offset onsets. m4_dark_vessel generates the synthetic Baltic scenes. Every output describes data consistency, never intent
Oracle
Synthetic scenes from the engine's own generator (about 2 000 vessels per seed in the Baltic), against bars pre-registered in docs/design/evidence/m4-dark-vessel-ais-integrity/. v1 (seeds 1000-1009, 20 000 vessels, ea8557adb): NOT MET on B2, dark recall 0.848 (bar 0.95; precision 0.999 meets its 0.80 part), and B4, offset recall 0.807 at 5 km (bar 0.85) and 0.846, 0.825 and 0.871 at 10, 20 and 50 km (bar 0.98 each); MET on B1 (false alarms 0.0050, bar 0.01), B3 gaps (600 of 600), B5 identity duplicates (600 of 600), B6 determinism on Linux and B7 additivity. v2 (seeds 2000-2009, 165db008b): NOT MET on B2, dark recall 0.877 (bar 0.95; precision 0.897), and B4, offset recall 0.396 at 5 km (bar 0.85) and 0.689, 0.650 and 0.646 at 10, 20 and 50 km (bar 0.98 each), lower than v1; MET on B1 (0.00006), B3 and B5 (1.000), B6 on Linux and B7; 141 offset vessels are decided DARK under v2. v3 (a turn model for dark tracks, a global assignment and collision-tolerant offset onsets, M4Params::v3; signed 10 November 2026; seeds 3000-3009, db06dc2ce): MET on B1 (false alarms 0.00075), B2 (dark recall 0.994, 1 392 of 1 400; precision 0.867), B3 and B5 (1.000), B6 on Linux and B7; NOT MET on B4, offset recall 0.489 at 5 km (bar 0.85) and 0.646, 0.675 and 0.661 at 10, 20 and 50 km (bar 0.98 each); 206 offset vessels are decided DARK and 501 NORMAL. M4 stays NOT MET (B4). WHAT IS NOT CHECKED: real vessels, real RF fixes or real AIS reception; a flag says nothing about intent (gaps and offsets have ordinary causes)
Tests
tests/m4_generator.rs (seed 0 byte-identical twice and equal to its pin); tests/m4_verdicts.rs (the recorded v1 and v2 verdicts, misses included; seeds 1000 and 2000 decided twice give the same decision and the…
ModelledInterference-source location from orbit, from a ground network, and from both combinedm5_hybrid
Capability
m5_hybrid: in a synthetic Baltic scenario, a three-satellite formation's fix of a surface or elevated interference source from its pass (link budget and detection probability per epoch), a coastal ground layer by power differences (locate_pd), and a hybrid that fuses both, with the information bound of each layer; v2 (m5_hybrid::run_v2) starts the space solver from a coarse search over the sea points and gives the ground layer 30 reports per site per draw
Oracle
Bars pre-registered in docs/design/evidence/m5-jammers-from-orbit-hybrid/, on a synthetic scenario (the scored class E30: 200 placements x 20 draws = 4 000 draws). v1 (ac44ae420): NOT MET on B1 to B3: success space 0.630, ground 0.000, hybrid 0.630 (bar: hybrid at least 0.680); medians hybrid 56.08 m, space 56.07 m (bar 44.86 m); p95 hybrid 359 km against ground 263 km (bar: at most the smaller). v2 (seeds 4600 and 4601, 4a09a6c9e): NOT MET on B1 to B3: success space 0.990, ground 0.00025, hybrid 0.990 (bar 1.040, which no estimator can reach); medians hybrid 7.352 m, space 7.351 m (bar 5.881 m; common set 262 draws); p95 hybrid 71.300 m against space 71.297 m (missed by 3.5 mm). MET in both: B4, the link budget and detection against an independent numpy recomputation (xval/m5-jammers-from-orbit/oracle.py) within 3e-14 dB and 7e-21 in probability at P1 to P10; B5, the hybrid bound never larger than a finite layer bound (0 violations in 600 cases); B6 determinism on Linux. The hybrid equals the space layer to millimetres; the ground layer as shipped does not locate (B3a's synthetic-gate miss carried over). WHAT IS NOT CHECKED: any real source, satellite or receiver
Tests
tests/m5_b4.rs (link budget and detection against the committed numpy oracle output); tests/m5_verdicts.rs (the recorded v1 and v2 E30 verdicts, misses included; the B5 bounds; two-placement runs give the pinned bytes)
ModelledMonitoring-network planner: detection and TDOA location-accuracy maps, and station placementplanner
Capability
planner (kshana planner map|place): for an interference source of a stated power at every cell of a site grid, each station's detection probability, the network's, the probability that enough stations detect to locate it, and the TDOA DRMS and 95 % ellipse the detecting stations bound (free space); planner::place chooses m stations from candidates, exhaustively up to 10 000 subsets, otherwise greedy followed by a single-exchange search (v1's greedy alone kept by name); v2 flags ill-conditioned cells
Oracle
An independent numpy and standard-library oracle (xval/planner/oracle.py), committed with its outputs before any planner code, under bars pre-registered in docs/design/evidence/b1a-planner/. v1: NOT MET on P1 (detection within 1.9e-13; 4 of the line layout's 619 locatable cells, collinear with every station, miss the 1e-9 relative DRMS bar by 1.3e-9 to 2.6e-9), P2 (52 of 54 barred cells within 5 %; the square's cell 1293 at 1.097 and the circle's at 0.717) and P4's greedy part (0.2754 against the best of 1 000 random subsets, 0.2790; the exhaustive part chose the oracle's optimum, objective 0.1689); P3, P5 (the same bytes on the Linux CLI, Python, WebAssembly and on Linux, macOS and Windows) and P6 MET. v2 (new layouts, 98ff742b4): P1 MET (1 662 locatable cells within 1e-9 relative, none with a condition number above 1e8), P2 MET (46 of 60 drawn cells in the regime, all within 5 %), P3 and P6 MET, P5 MET on Linux (the cross-platform half is met when the reproducibility matrix runs the test on macOS and Windows); P4 NOT MET: on 40 candidates with m = 7, greedy and exchange give 0.350982 against the best of 1 000 random subsets, 0.353361 (greedy alone 0.324212, no bar). v3 (greedy, double exchange and 8 restarts, kept by name; signed 10 November 2026; a new 48-candidate instance with m = 8, 060ee2a07): P4 MET, objective 0.440809 (median DRMS 19.23 m) against the best of 1 000 random subsets, 0.410470; P1 to P3 and P6 MET, re-run unchanged; P5 MET on Linux (the macOS and Windows half is pending, as for v2; the v3 placement's consistency check did not complete when the container restarted, and the printed values are pinned). B1a part 1 is MET on Linux; B1b starts when P5's cross-platform half is green. WHAT IS NOT CHECKED: any real site; the maps are free-space predictions and the DRMS is a lower bound
Tests
tests/planner.rs (v1 P1 to P5, misses included); tests/planner_v2.rs (v2 P1 to P5; the pentagon map's SHA-256 pinned); tests/planner_v3.rs (the v3 P4 values pinned, ignored by default, about 1.5 h)
ModelledArea trust state from several reference stationsarea_trust
Capability
area_trust (kshana area-trust, Python and WebAssembly): each station's epochs on a 1 s grid (Stale after 5 s with no epoch) combine into one area state with k = max(2, ceil(N/3)): UNKNOWN with fewer than 2 live stations, UNTRUSTED with at least k untrusted, DEGRADED with at least k degraded or worse or any untrusted (v2, selected by the stations document's area_rule field: at least m untrusted, m = 2 for up to 6 stations and 1 above), else TRUSTED; a worse state is published at once, a better one after 10 s. v3 (a kshana.area-measurements/1 document with per-satellite inputs, area_trust::v3): per epoch, when at least k stations exceed H_s, the one satellite whose removal brings every alarmed station under its H_s is flagged SATELLITE_FAULT and every station's fix is recomputed without it; loss of lock alone no longer degrades a station; the aggregation is v2's. area_trust::synth and area_trust::v3::fault_replay build seeded replays with injected events
Oracle
Bars pre-registered in docs/design/evidence/b5-area-trust/. T1 against an independent Python standard-library recomputation of the rule (xval/area-trust/oracle.py and oracle_v2.py); T4 on real IGS observation files (SHA-256 values in xval/area-trust/MANIFEST.tsv and MANIFEST_v2.tsv). v1: T1 MET (43 117 grid seconds of seven seeded replays, 4 to 50 stations, 520 events; the oracle's bytes); T2 MET (all 267 events with at least k stations in alarm raised the expected alert; none of the 253 with fewer raised UNTRUSTED); T3 MET (220 injected events at 100 stations, 1 Hz, over MQTT: webhook and MQTT p95 1.004 s, maximum 1.015 s; bars 5 s and 10 s); T5 MET for the files and surfaces built; T4 NOT MET: 23 alerts to DEGRADED over 13 area-days (1.77 per area-day; bar 1), none to UNTRUSTED (bar 0.1), Madrid (MADR, CEBR, YEBE, VILL), 2025-01-07 to 01-19. v2 (a two-station floor for areas of up to 6 stations, a068a3eb9): T1 MET (43 165 grid seconds, the v2 oracle's bytes), T2 MET, T3 MET (p95 1.004 s, maximum 1.013 s), T5 MET; T4 NOT MET: 16 alerts to DEGRADED or worse over 13 area-days (1.23 per area-day; bar 1), 9 of them to UNTRUSTED (0.69 per area-day; bar 0.1), Goddard (GODS, GODE, GODN, USN7), 2025-07-25 to 08-06; no alert came from one station alone. v3 (a common-mode satellite-fault screen across the area's fixes, loss of lock alone no longer degrading, signed 10 November 2026; the diagnosis found unhealthy broadcast records in use and a stale-record satellite fault, and the engine now skips unhealthy GPS records): T1 MET (520 epochs and 520 grid seconds of four satellite-fault replays equal oracle_v3.py's bytes; errors within 1e-6 m), T2 NOT MET (123 of 127 injected common faults flagged and cleared, 4 not where the remaining geometry is weak; 2 of 67 two-satellite epochs explained by one), T3 MET (220 events at 100 stations over MQTT, webhook and MQTT p95 1.003 s and 1.004 s, maximum 1.006 s; the screen is not in its path), T4 MET (1 alert to DEGRADED or worse and 1 to UNTRUSTED over 26 area-days, 0.038 each; Goddard and Wettzell, 18 to 30 May 2026, the pre-registration's 2025 window having no quiet fortnight, disclosed before any fetch), T5 MET (command line, Python and WebAssembly); B5 v3 NOT MET (T2). WHAT IS NOT CHECKED: AGC, the jamming indicator, the clock monitor and OSNMA on the real-data days (not in the files); Galileo and BeiDou health; any real interference event
Tests
tests/area_trust.rs (the v1 and v2 timelines equal the oracles' bytes every second; T2 alerts; boundary rules; the live rule equals the timeline; the T4 misses and T3 v2 recorded); tests/area_trust_igs.rs (T4 and T3 on…
ModelledInterference-source location from power differences across fixed monitoring stationslocate_pd
Capability
locate_pd: each station's C/N0 drop (and AGC rise, where its dB-per-unit scale is stated) is turned into an interference power at the antenna; a log-distance model with the reference power profiled out, the path-loss exponent and source height marginalised and censored terms for stations that see nothing gives a posterior on a grid around the network, with its 95 % region and centroid. v2 averages each station's power over the session with one shadowing term per station, no tempering, on the stations' hull buffered by 10 km; v1 is kept as recorded
Oracle
Synthetic networks drawn from a model that differs from the locator's on purpose, with an independent numpy recomputation (xval/locator-power-difference/oracle.py), under bars pre-registered in docs/design/evidence/b3a-power-difference-locator/. v1 gate (seeds 1001-2000, tau 1/60, f6b4a3336): S1 coverage MET (0.976; bar 0.90); S2 NOT MET (median 95 % region 52.5 % of the search domain; bar 25 %); S3 NOT MET (median centroid error 2.55 x the median nearest-neighbour station spacing; bar 0.5 x). v2 gate (session-level shadowing, seeds 4001-5000, 53ca11af5): NOT MET on S1 (0.890; bar 0.90; the 34 of 1 000 sessions not located count against it), S2 (26.1 %; bar 25 %) and S3 (1.45 x; bar 0.5 x); S4 determinism platforms only (Linux debug against release; the CI job b3a-v2-determinism runs it on Linux, macOS and Windows; the CLI, Python and WebAssembly half is not assessable, the locator has none of those surfaces); the numpy recomputation agrees on seeds 4001-4020 (centroids within 2.5e-10 m). By the pre-registration the real run on Jammertest 2027 is not made with this method. WHAT IS NOT CHECKED: any real interference; the C/N0 and AGC monitors themselves
Tests
tests/locator_power_difference_synthetic.rs (the v1 gate rows reproduce; the recorded v1 verdict); tests/locator_power_difference_v2.rs (12 v2 gate seeds reproduce bit for bit, debug against release; the recorded v2…
ModelledInterference-source location by time difference of arrival across fixed monitoring stationslocate_tdoa
Capability
locate_tdoa: from 10 ms snapshots at 4 Msps time-tagged by each station's clock, the station with the largest measured jammer-to-noise ratio is the reference; each other station's delay is the cross-correlation peak refined by a parabola, with a sigma from the Knapp-Carter bound for the measured SNRs and the reference's mean-square bandwidth plus both clocks (the pairs share the reference clock's covariance); the source is located on a spherical Earth at a known height by a grid search over the stations' hull buffered by 10 km and Gauss-Newton, with a 95 % ellipse and a chi-square test of the residuals at 0.999 that gives not located. locate_tdoa::session draws the synthetic sessions (CW, chirp and wideband noise jammers, two-ray ground reflection with 6 dB shadowing, 20 ns station clocks)
Oracle
Synthetic sessions under bars pre-registered in docs/design/evidence/b3b-tdoa-locator/ (B3a's bars), with an independent numpy recomputation of the locator (xval/tdoa-locator/oracle.py). Gate (seeds 6001-7000, 1 000 sessions, eed00fd84): NOT MET on T1 (coverage 0.044; bar 0.90), T2 (median area share 1.000; bar 0.25) and T3 (3.35 x the median nearest-neighbour spacing; bar 0.5); 872 sessions not located (CW 334 of 334, chirp 324 of 333, noise 214 of 333). The parabolic refinement misplaces a one-sample-wide correlation peak by up to 125 ns against a stated sigma of about 28 ns, so the chi-square test rejects most fits; the sawtooth chirp's correlation repeats inside the lag window; a CW has no bandwidth. The numpy recomputation agrees on 19 of seeds 6001-6020 within the registered tolerances; on seed 6002 Gauss-Newton diverges in both and the scored outputs agree. By the pre-registration no real run is made with this method. WHAT IS NOT CHECKED: any real jammer, station or clock
Tests
tests/locator_tdoa.rs (the recorded gate verdict, the misses included; seeds 6001-6020 against the numpy oracle; the whole gate reproduces its recorded bytes, ignored by default and run on Linux, macOS and Windows by…
ModelledTEXBAT spoofing-detection benchmark: per-window baseline IQ scores and power-and-distortion scoresbench::texbat; bench::cli
Capability
bench::texbat (kshana bench texbat-windows and texbat-pd): streams one TEXBAT recording (complex int16, 25 Msps), refuses a file whose SHA-256 or size differs from the pinned manifest and then writes nothing, acquires PRNs 1 to 32 once and tracks them through the recording, and writes per 1 s window the early-minus-late imbalance S, the prompt-power fluctuation P and the window score against synthetic thresholds (the 99th percentiles over the 61 of 64 synthetic windows that acquire); texbat-pd adds the in-band power rise over the recording's own windows 0 to 59 (tau 1.0 dB). Writes scores only, never IQ samples
Oracle
TEXBAT (Humphreys et al., ION GNSS 2012): two clean and eight spoofing recordings, spoofing windows counted from 110 s, under bars pre-registered in docs/design/evidence/b4a-benchmark-classes/PREREGISTRATION-E5-v2.md (E5 v2: window AUC of spoofing at least 0.70 over all ten recordings) and PREREGISTRATION-E5-PD.md (E5-PD: the same bar on windows 60 and later). Scored once on 11 November 2026 (kshana bench texbat-score over the ten committed score files, xval/bench-e/texbat-v2/E5-SCORE.json, 571cb9d19); every scoring rule, the thresholds, the onset and the window rules were committed before the first TEXBAT score file (9 November). E5 v2 NOT MET: AUC 0.6941 (2 729 spoofing and 922 nominal windows; bar 0.70). E5-PD MET: AUC 0.7410 (bar 0.70). Reported without a bar: recording-level AUC 0.8125 (E5 v2); per spoofing recording, E5-PD: ds1 0.466, ds2 1.000, ds3 0.418, ds4 0.821, ds5 1.000, ds6 0.584, ds7 0.897, ds8 0.750 (three of the eight below 0.6). An independent Python standard-library recomputation (xval/bench-e/texbat-v2/oracle_e5.py, exact pairwise AUC) agrees to 1e-12. The thresholds come from synthetic data only; the recordings are not redistributed (hashes and score files only). The detector stays MODELLED. WHAT IS NOT CHECKED: real receivers in the field; attacks outside TEXBAT
Tests
tests/texbat_e5v2.rs (the score rules and classes; a file that does not match its pin writes nothing; a power-and-distortion run refuses a windows file of another recording; the synthetic thresholds, ignored by…
ModelledRTKLIB .pos solution reader for receiver-trustreceiver_trust::machines::rtklib (read_rtklib_pos); receiver_trust::LogFormat::RtklibPos
Capability
Reads an RTKLIB solution file (calendar or GPS week/TOW time, latitude/longitude or ECEF position, satellites used) into the receiver-trust timeline (receiver_trust::machines::rtklib).
Oracle
None external: no RTKLIB-produced file was read, and the column layout is from the RTKLIB manual. Checked on self-written files only; the ECEF conversion reuses frames::ecef_to_geodetic, which has its own rows
ModelledNovAtel OEM7 ASCII and binary logs (BESTPOS, RANGE, RXSTATUS) for receiver-trustreceiver_trust::machines::novatel (read_novatel,novatel_crc32); receiver_trust::LogFormat::Novatel
Capability
Reads BESTPOS (position), RANGE (C/No per signal) and RXSTATUS (the receiver's jammer flag, as the epoch's self-report) from ASCII and binary OEM7 logs, CRC-checked, read-only (receiver_trust::machines::novatel).
Oracle
None external as a pass bar. The field layouts and CRC follow the public OEM7 documentation (docs.novatel.com); the CRC routine was checked once against the CRC the documentation states for its own BESTPOSA example (cited, not quoted). Fixtures are self-written; no log from a real receiver was read
Tests
receiver_trust::machines::novatel::tests (pinned CRC, ASCII and binary records, bad CRC skipped, hostile input); tests/receiver_trust_machines.rs::novatel_jammer_flag_reaches_the_receiver_self_report
ModelledRTCM 3 MSM4 and MSM7 carrier-to-noise reader for receiver-trustreceiver_trust::machines::rtcm3 (read_rtcm3,crc24q); receiver_trust::LogFormat::Rtcm3
Capability
Reads the C/N0 of every signal cell of RTCM 3 MSM4 and MSM7 messages (GPS, GLONASS, Galileo, SBAS, QZSS, BeiDou, NavIC) from a frame stream, CRC-24Q checked (receiver_trust::machines::rtcm3). Layouts from the BSD-licensed pyrtcm tables, not from the RTCM standard text.
Oracle
ReferenceImpl in a one-off check, not a test in the suite: pyrtcm 1.2.0's decoder reads the same self-written frames to the same C/N0 and epoch time. It shares its tables' source with this reader, so it checks the bit handling, not the layout. No capture from a real receiver or base station was read
Tests
receiver_trust::machines::rtcm3::tests (CRC-24Q catalogue check value, MSM4/MSM7 C/N0 and epoch merging, GLONASS time and SBAS id mapping, bad CRC and other messages)
ModelledGNSS-SDR monitor adapter for receiver-trustreceiver_trust::machines::gnss_sdr (read_gnss_sdr_monitor); receiver_trust::LogFormat::GnssSdr
Capability
Reads GNSS-SDR Monitor (Gnss_Synchro) and PVT-monitor messages, relayed to JSON lines with the proto field names, into per-channel C/N0, fix, satellites used, HDOP and sky (receiver_trust::machines::gnss_sdr; integrations/gnss-sdr/relay.py).
Oracle
None external: field names come from the GNSS-SDR documentation; neither the reader nor relay.py was run against a live GNSS-SDR or a real capture
ModelledIMU log ingest, gyro heading input and Allan fit of hybrid sensor parametersreceiver_trust::machines::imu,imu_fit; surface::imu_fit; scenario [log.imu]
Capability
Reads an IMU CSV (documented columns, or the EuRoC layout); integrates the yaw rate into the heading the heading-course monitor reads ([log.imu]); and fits the accelerometer and gyro noise parameters of the hybrid and fusion scenarios from a stationary log by overlapping Allan deviation (imu-fit).
Oracle
None external: the fit is checked on synthetic white-noise logs with known coefficients, the heading on constant yaw-rate logs. No real IMU log was fitted, and q_aa is not fitted
Tests
receiver_trust::machines::imu::tests; receiver_trust::machines::imu_fit::tests (known white-noise coefficients recovered within 10 %, refusals, TOML parses as AccelCfg); tests/receiver_trust_machines.rs (heading…
ModelledHome Assistant MQTT discovery and state exporter for trust telemetrytelemetry::mqtt (discovery,state_json,connect_packet,publish_packet,MqttSink); surface::TelemetryFormat::Mqtt
Capability
A JSON state document and Home Assistant MQTT discovery messages for the trust stream, printed as text on every surface and published over MQTT 3.1.1 QoS 0 by the CLI (telemetry::mqtt; trust-telemetry --mqtt-tcp, --print-mqtt, format mqtt).
Oracle
None external: the packets were read by a loopback listener written for the test, not by a real broker, and Home Assistant was not run
Tests
telemetry::mqtt::tests (packet bytes, loopback broker, dead broker counted); tests/receiver_trust_machines.rs::trust_telemetry_mqtt_text_has_discovery_then_state
ModelledROS 2 receiver-trust package and gnss_lib_py bridgeintegrations/ros2/kshana_trust (bridge.py,trust_node.py); integrations/gnss_lib_py/kshana_gnss_lib_py.py
Capability
A ROS-free bridge that writes fix and satellite messages as NMEA and scores them through the Python binding (integrations/ros2/kshana_trust), and a gnss_lib_py NavData to receiver-trust bridge through the rtklib-pos format (integrations/gnss_lib_py).
Oracle
None external: the ROS node was not run on a live ROS 2 graph, and the ublox_ubx_msgs field names were not checked against a running driver. gnss_lib_py (MIT) was run for the bridge test
Tests
tests/receiver_trust_machines.rs (rtklib_pos_runs_as_a_receiver_log: the format the gnss_lib_py bridge writes, run through receiver-trust); the Python tests of the bridges are…
Nothing matches that filter.
Generated from the engine's verification matrix and pinned by a test. Scroll the list, or open a row for its capability, oracle and tests. The full table, every column at once, is in the docs: Capability reference.
ProNeed this ledger as an audit-grade, reproducible evidence pack for a review board? Kshana Pro packages it.
04Reproducible by default
Rerun it. Get the same numbers.
Free core A result is reproducible bit for bit from scenario, seed and engine version. Nothing here needs a licence, an account or us.
The stamp every result.json carriesRecorded run
This one is from the recorded clock-holdover run. Open it in Kshana Studio, or rerun it yourself:
Run the same scenario.It writes result.json, chart.svg, report.html and table.csv next to the scenario.
Compare the stamp.The scenario hash is a SHA-256 (256-bit Secure Hash Algorithm) digest over the canonical scenario: seed, thresholds, model parameters. Change any input and the hash changes. Every chart footer carries its first 12 hex characters.
Let the engine check itself.On every build, tests confirm that the published figures still reproduce (tests/published_figures_still_reproduce.rs) and that every ledger row names a test and cites evidence that exists (tests/verification_rows_name_a_test_that_exists.rs, tests/verification_rows_cite_evidence_that_exists.rs).
05Standards grid
Speaks the formats you already file.
18 standards and formats the engine implements. A green mark means it is checked against real data, a published example or an independent reference implementation.
TLE / SGP4AIAA 2006-6753CheckedTLE two-line element set · SGP4 Simplified General Perturbations 4, an orbit model · AIAA American Institute of Aeronautics and Astronautics
RINEX 3real IGS data → surveyed coordinateCheckedRINEX Receiver Independent Exchange Format · IGS International GNSS Service
SP3-c / dfit to real ESA/ESOC precise orbitsCheckedSP3 Standard Product 3, a precise-orbit format · ESA European Space Agency · ESOC European Space Operations Centre
CCSDS OEM / OMMparses the CCSDS 502.0-B-3 Blue Book OEM exampleCheckedCCSDS Consultative Committee for Space Data Systems · OEM orbit ephemeris message · OMM orbit mean-elements message
IS-GPS-200broadcast ephemeris → real-data PVTCheckedIS-GPS-200 GPS interface specification for the L1 and L2 signals · GPS Global Positioning System · PVT position, velocity and time
IEEE 1139ADEV/MDEV/TDEV/HDEV vs NIST SP 1065 reference deviationsCheckedIEEE Institute of Electrical and Electronics Engineers · ADEV Allan deviation · MDEV modified Allan deviation · TDEV time deviation · HDEV Hadamard deviation · NIST National Institute of Standards and Technology
SBAS / DO-229EDO-229E HPL/VPL vs RTKLIB SBAS-PL fork on real EGNOS data · L1/L5 iono-freeCheckedSBAS satellite-based augmentation system · DO-229E RTCA minimum operational performance standards for satellite-based augmentation system receivers, revision E · HPL horizontal protection level · VPL vertical protection level · RTKLIB an open-source GNSS positioning library · EGNOS European Geostationary Navigation Overlay Service
IONEX / KlobucharKlobuchar L1 delay vs RTKLIB ionmodel · IONEX TEC-map reader lenient, no external oracle yetCheckedIONEX Ionosphere Map Exchange format · RTKLIB an open-source GNSS positioning library · TEC total electron content
IGRF-14vs official BGS IGRF-14 valuesCheckedIGRF-14 International Geomagnetic Reference Field, 14th generation · BGS British Geological Survey
CCSDS 133.0primary header vs CCSDS 133.0-B-2 + spacepackets-py vectorsCheckedCCSDS Consultative Committee for Space Data Systems
CCSDS TDMparses the CCSDS 503.0-B-2 Blue Book TDM exampleCheckedCCSDS Consultative Committee for Space Data Systems · TDM tracking data message
CCSDS 401 / DSNlink equation vs published DESCANSO/JPL Galileo DCTCheckedCCSDS Consultative Committee for Space Data Systems · DSN Deep Space Network · DESCANSO Deep Space Communications and Navigation Systems Center of Excellence · JPL Jet Propulsion Laboratory · DCT design control table, the link-budget worksheet
IMO GNSS-interference guidance (ships)IMO guidance for ships, mapped by clause numberImplementedGNSS Global Navigation Satellite System
EU NIS2, Article 21risk-management measures, mapped by clause numberImplemented
EN 16803space-based PNT in transport, mapped by clause numberImplementedPNT positioning, navigation and timing
US DHS resilient-PNT frameworkresilient-PNT conformance framework, mapped by clause numberImplementedDHS United States Department of Homeland Security · PNT positioning, navigation and timing
EASA GNSS-interference guidance (aviation)EASA guidance for aviation, mapped by clause numberImplementedGNSS Global Navigation Satellite System
Nothing matches that filter.
When a review needs more
Same evidence, packaged or done for you.
The engine and everything above stay free. Two paid routes build on them, under contract.
Pro
Kshana Pro: a mission dossier and an evidence pack
Kshana Pro packages the evidence a review or procurement board asks for, reproducibly. What it builds on: this ledger, where every capability names its oracle and its test, and the stamp every result carries. In one command, a mission dossier checks every requirement against real runs and writes a verification matrix with its open items and a PDF (Portable Document Format).
A custom study answers your own question with the same labels: every figure Validated or Modelled, not a certification. Export-controlled work is handled under clearance and NDA (non-disclosure agreement).
kshana-bench fixes the tasks, splits and metrics for GNSS interference detection and timing holdover, with baselines whose bars were written down before the first scored run. A score describes a log or a corpus, never a product.
6 papers on arXiv, the open preprint archive, built on the open engine. Each opens to a figure rerun on the engine, its findings, the command and its DOI (Digital Object Identifier).
Lunar navigation · arXiv:2610.23871
Cislunar navigation is an observability problem: arc-length thresholds, Gramian conditioning, and the Doppler lever
In cislunar space dilution of precision does not apply: range-only links leave the velocity unobservable at any instant, so requirements belong in arc-length, conditioning and measurement-type budgets.
FigureModelledEngine v0.41.0 · cislunar-observability.toml · deterministic, no seedOpen in Kshana StudioKey findings
In the engine run, range-only geometry at one instant has rank 2 of 4, so a dilution of precision is undefined; the full state becomes observable after a 2.09 h arc.
Adding range rate lifts the instantaneous rank from 2 to 4: a Doppler-capable link makes a snapshot solution exist where a range-only link cannot.
Beyond the threshold the conditioning improves with arc length, so requirements read better as arc-length, conditioning and measurement-type budgets.
How Kshana reproduces it
The cislunar-observability kind differential-corrects four planar distant retrograde orbits, assembles the range and range-rate Jacobians through the three-body state-transition matrix, and reports the observable rank over a growing tracking arc, the observability Gramian's condition number, and a square-root information filter that recomputes the same quantity by a second route.
The figure. The bundled scenario: four differential-corrected planar distant retrograde orbits, tracked by inter-satellite ranging over a 6 h arc sampled every 0.26 h.
Reproduce it
Runs the scenario kind cislunar-observability (scenarios/cislunar-observability.toml) with the installed engine.
@misc{baweja2026cislunar,
title = "{Cislunar navigation is an observability problem: arc-length thresholds, Gramian conditioning, and the Doppler lever}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2610.23871},
archivePrefix = {arXiv},
primaryClass = {eess.SP},
doi = {10.48550/arXiv.2610.23871},
url = {https://arxiv.org/abs/2610.23871}
}
Lunar navigation · arXiv:2607.06212
The Cost of Lunar South-Polar Geometry, and Surface Beacons as the Efficient Fix: A Dilution-of-Precision Analysis
Planned orbits bunch overhead for a lunar south-pole user; a few ranging beacons on nearby high ground fix the geometry more cheaply than more satellites.
FigureModelledEngine v0.41.0 · lunar-beacon.toml · deterministic, no seedOpen in Kshana StudioKey findings
In the bundled run, the one beacon that clears the horizon cuts the geometric dilution of precision from 11.66 to 4.998; a 24-satellite constellation instead reaches 2.893.
The satellites a south-polar user sees bunch into a small patch overhead, so the geometry is limited by their spread, not their number.
A few beacons on elevated terrain reach geometry that an orbit-only design reaches only with a much larger fleet.
How Kshana reproduces it
The lunar-beacon kind builds the ranging geometry of an orbital constellation plus the surveyed surface beacons that clear the airless horizon, on the dilution-of-precision kernel checked against an independent library, and reports each configuration's dilution of precision and its accuracy in metres.
The figure. One epoch of the bundled scenario, its default geometry: the satellites alone, the satellites plus the beacons that clear the horizon, and a 24-satellite constellation instead.
Reproduce it
Runs the scenario kind lunar-beacon (scenarios/lunar-beacon.toml) with the installed engine.
@misc{baweja2026cost,
title = "{The Cost of Lunar South-Polar Geometry, and Surface Beacons as the Efficient Fix: A Dilution-of-Precision Analysis}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2607.06212},
archivePrefix = {arXiv},
primaryClass = {astro-ph.EP},
doi = {10.48550/arXiv.2607.06212},
url = {https://arxiv.org/abs/2607.06212}
}
Lunar navigation · arXiv:2607.02566
Earth-baseline VLBI restores the observability of a lunar surface station in joint orbit-and-clock determination
Lunar-network ranging fixes only relative geometry; an Earth-baseline very-long-baseline interferometry delay ties a surface station to the Earth frame, even with few satellites. VLBI very long baseline interferometry
FigureModelledEngine v0.41.0 · lunar-joint-od-clock.toml · seed 42Open in Kshana Studio
Recorded on macOS as published. A fresh Linux x86-64 run gives 2184 m / 3.537 m (solver iterate host dependence, issue #36 row 6).
Key findings
Ranging inside the lunar network leaves the cluster free to shift and rotate unseen; the clocks stay observable, so the defect is purely positional.
In the engine run (3 satellites), the Fisher information without the interferometry legs has rank 15 of 16; with them it has full rank and a Cramér–Rao bound of 20.1 m.
With a rich constellation the interferometry tie only sharpens the bound; with a sparse one it restores the station's absolute position.
How Kshana reproduces it
The lunar-joint-od-clock kind fuses Earth-baseline interferometry delays, lunar-local ranges and inter-satellite ranges in one batch least-squares fit, runs it with and without the interferometry legs on the same seed, and reports the rank of the Fisher information, the datum defect and the Cramér–Rao bound.
The figure. The bundled simulated network: 3 lunar satellites, 6 Earth stations, one surface station, the same seed and truth for both fits.
Reproduce it
Runs the scenario kind lunar-joint-od-clock (scenarios/lunar-joint-od-clock.toml) with the installed engine.
@misc{baweja2026earth,
title = "{Earth-baseline VLBI restores the observability of a lunar surface station in joint orbit-and-clock determination}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2607.02566},
archivePrefix = {arXiv},
primaryClass = {eess.SP},
doi = {10.48550/arXiv.2607.02566},
url = {https://arxiv.org/abs/2607.02566}
}
Timing integrity · arXiv:2606.24210
A Conditional Timing Protection Level: Holdover-Limited Undetected Time Error Under GNSS Spoofing
A recorded spoof shows a timing receiver serving time far more wrong than it reports; no unconditional bound exists, so the paper bounds the undetected error conditionally. GNSS Global Navigation Satellite System
FigureModelledEngine v0.41.0 · examples/tpl_jammertest.rs · deterministic, no seedKey findings
A public recorded spoof pulled the receiver's served time by about 1.01 ms while it reported at most 51 ns.
Against a slow enough ramp no finite unconditional bound exists, so the bound is conditional on an independent cross-satellite check detecting the attack.
The engine's Timing Protection Level is 114 ns at 1 s detection latency and 458 ns at 60 s.
How Kshana reproduces it
The Timing Protection Level module adds a monitor's detection floor to the oscillator's coast over the detection latency. The example calibrates it with the scalars recovered from the public JammerTest 2024 recording; the raw recording is not shipped.
The figure. The example's printout: the conditional Timing Protection Level at five detection latencies, and the two scalars it recovered from the public JammerTest 2024 recording.
Reproduce it
Regenerates the study artifact from examples/tpl_jammertest.rs, in a clone of the repository.
Cite it
BibTeX · baweja2026conditional
@misc{baweja2026conditional,
title = "{A Conditional Timing Protection Level: Holdover-Limited Undetected Time Error Under GNSS Spoofing}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2606.24210},
archivePrefix = {arXiv},
primaryClass = {eess.SP},
doi = {10.48550/arXiv.2606.24210},
url = {https://arxiv.org/abs/2606.24210}
}
Resilience scoring · arXiv:2607.05415
How Stable Is a PNT Resilience Score? Decision-Instability of Single-Number Resilience Ratings under Framework-Aligned Weighting
Single-number resilience ratings for positioning, navigation and timing are stable only where one design dominates; weighting, threat and declared techniques can flip them. PNT positioning, navigation and timing
Re-weighting alone changes the top-ranked architecture in 22% of draws under nominal conditions, and in at most 1.25% under each active threat.
A weakest-link maturity level depends on the threat assumed, not only on the architecture.
Because the composite rewards declared techniques, a design that declares more can outscore a more resilient one.
How Kshana reproduces it
The resilience module scores seven reference architectures across the seven framework categories and five threats, draws the category weights at random from a Dirichlet distribution and counts how often the top rank changes. The example writes the whole study as one deterministic file.
The figure. 7 reference architectures, 2,000 random category weightings per threat.
Reproduce it
Regenerates the study artifact from examples/resilience_report.rs, in a clone of the repository.
Cite it
BibTeX · baweja2026stable
@misc{baweja2026stable,
title = "{How Stable Is a PNT Resilience Score? Decision-Instability of Single-Number Resilience Ratings under Framework-Aligned Weighting}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2607.05415},
archivePrefix = {arXiv},
primaryClass = {cs.CR},
doi = {10.48550/arXiv.2607.05415},
url = {https://arxiv.org/abs/2607.05415}
}
Radio-frequency interference · arXiv:2606.22054 · version 2
Anticipating the Optimism Gap: Predicting Distribution-Shift Degradation of RF-Impairment Detectors from In-Distribution Statistics
How much an interference detector loses under shifted conditions can be predicted from its own in-distribution scores; synthetic results, checked on open field recordings. RF radio frequency
The gap between in-distribution and shifted area under the curve grows as the shift deepens, and depends on how many observables a detector uses rather than on whether it is learned.
In the engine run, a ridge model on in-distribution score statistics predicts the gap for a detector it has not seen (R² 0.46) and for an interference class it has not seen (R² 0.33).
The headline findings are synthetic; the paper then checks them on open field recordings.
How Kshana reproduces it
The impairment study generates a synthetic, parameter-grounded corpus for four interference classes, scores thirteen detectors at four shift severities over five seeds, and fits the gap predictor with leave-one-out cross-validation and a permutation test.
The figure. 260 detector, class and severity cells; each prediction is made for a detector left out of the fit.
Reproduce it
Regenerates the study artifact from examples/optimism_study.rs, in a clone of the repository.
Cite it
BibTeX · baweja2026anticipating
@misc{baweja2026anticipating,
title = "{Anticipating the Optimism Gap: Predicting Distribution-Shift Degradation of RF-Impairment Detectors from In-Distribution Statistics}",
author = {Baweja, Chakshu},
year = {2026},
eprint = {2606.22054},
archivePrefix = {arXiv},
primaryClass = {eess.SP},
doi = {10.48550/arXiv.2606.22054},
url = {https://arxiv.org/abs/2606.22054}
}
Cite the engine
Every release is archived on Zenodo with a citable DOI, and a CITATION.cff ships in the repository. Please cite the engine version, the scenario and the seed.
Baweja, C. (2026). Kshana — a PNT-resilience simulator with quantum-sensor performance models (Version 0.41.0). Ashforde OÜ. https://doi.org/10.5281/zenodo.20528627
BibTeX for the engine
BibTeX
@software{kshana,
title = {Kshana — a PNT-resilience simulator with quantum-sensor performance models},
author = {Baweja, Chakshu},
version = {0.41.0},
year = {2026},
doi = {10.5281/zenodo.20528627},
url = {https://github.com/ashfordeOU/kshana},
license = {AGPL-3.0-only}
}