Where each capability is reachable

Every capability is available on the surfaces where it makes sense; where one cannot carry a capability, the cell says N/A and why. The same engine runs behind every surface: the Python, WebAssembly and MCP entry points for the 0.35 capabilities are thin wrappers over kshana::surface, which calls the code the command line runs.

A passage naming third-party AI clients is left out here; it is in docs/SURFACES.md on GitHub.

Studio#

Kshana Studio is the web app built on the WebAssembly build, so a Studio view exists where the WASM column has an entry point, and where it does not the cell points to a Command line card (the command to run, with its arguments), never to a result Studio computed. Pages (kshana.dev/studio/, client-side): Maritime trust (trust/), Training debrief (training/), Interference map (interference/) and the Tools page, whose sections are OSNMA (#osnma), Evidence packs (#evidence), Test-bench export (#bench), Compliance mapping (#compliance), the Command line tab (telemetry card, IQ, marine integrations: commands to run) and Use from your AI assistant (#ai). A — in the Studio column of the subcommand table is a command that is not one of the 0.35 capabilities Studio presents.

Why the N/A cells are N/A#

• Signal K and OpenCPN. They are consumers of the running live process and of its TCP gate, so they are command-line deployments; the plugin carries a how-to command and a skill that point to docs/MARINE-INTEGRATIONS.md and call no tool.

• A running process. receiver-trust live holds a stream open, can serve the gated stream to any number of TCP clients and runs until it is stopped. A Python call, a browser call, an MCP tool call and an IDE action each return when they finish and do not hold a socket open for others. Python, WASM and MCP carry the bounded form instead: an excerpt scored by the same engine, in the live schema (JSON lines 1.2: position and an advisory statement on every epoch), with no socket and no gate (the gate marks a fix invalid on a live stream, which only a process in the data path can do). Two behaviours belong to the process and are not in the excerpt tools: a stream replayed faster than real time through receiver-trust live without --replay is scored untrusted (its time disagrees with the host clock; an excerpt is read without a host clock), and --udp <port> and --listen tcp:<port> bind the loopback address unless an address is given.

• Trust telemetry. WASM carries trust_telemetry(input, format, host, input_kind, options_json): the same encoders, returning Prometheus text, CEF or LEEF syslog lines, or the OTLP JSON payload for a stream or a batch result, with no clock (syslog lines carry the RFC 5424 nil timestamp) and nothing sent. Python (trust_telemetry) and MCP (trust_telemetry, with the plugin command /kshana-trust-telemetry) carry it too. The rest of trust-telemetry serves Prometheus metrics, sends syslog and exports OpenTelemetry from a running stream, so it is a process by nature. How-to: pipe the live command into it, or replay a result file (kshana trust-telemetry --result session.result.json --print-metrics); see TRUST-TELEMETRY.md. An agent or a notebook that wants the per-epoch scores calls the batch or excerpt functions instead.

• Creating a pack in the browser. Signing needs a private key, which should not pass through a web page; the browser build verifies packs (evidence_verify) and the other surfaces create them. The MCP tool never returns or logs the signing seed, and without one it signs with a one-time key (integrity, not identity); for a key that matters, use the CLI. The same holds for streaming training NMEA to an address: Python, WASM and MCP return the generated text, and the CLI sends it.

• Study suites. --study <suite.toml> runs a manifest that names sibling scenario files on disk. The in-memory surfaces take scenario text, not a folder of files, so they run each member as a scenario instead (run, run_scenario, run_all).

• GNSS IQ in the browser, and recordings that are rewritten. A recording is gigabytes of samples, which a page cannot hold and the MCP protocol must not carry, so WASM has no IQ functions and the MCP tools read and write one work directory only (KSHANA_MCP_IQ_DIR, with a per-call sample budget). iq extract, convert and decimate exist to rewrite recordings on disk and stay command-line only. iq labfit and iq conditions carry text, so MCP takes them inline (a log that names a file is refused).

• The land download. It is the only command that opens a network connection, it needs the system curl and an explicit --allow-network, and it checks a pinned SHA-256. The browser has no curl, and an AI agent should not fetch data on a user's behalf; the other surfaces take the land file as text instead (land_geojson).

• Trace archives and directories. interference-map adsb on a readsb trace directory reads files from disk and is command-line only; the in-memory surfaces take CSV text.

Every command-line subcommand, and who covers it#

One row per subcommand or mode of kshana --help. yes names the entry point; no is a gap with its reason; N/A is a surface that cannot carry it for the reason given above or in the last column. JetBrains actions run the command line, so an N/A there states why no IDE action exists.

Utility functions#

Small entry points that are not a capability row, and where each lives:

Limits that differ by surface#

Evidence tiers and caveats on every surface#

• An evidence pack is a technical record of what the engine computed from a log under stated settings, with every hash and a signature over them. It is not a legal opinion, a finding of fact about any event or an approval. A pass of verify says the record is unchanged; with no trusted public key it proves only integrity against the key the pack names itself.
• The receiver-trust monitors and the trust score are MODELLED: thresholds are stated inputs and nothing asserts how they perform against real interference. The output is advisory; Kshana is not type-approved navigation equipment (IEC 61108, IEC 61162) and the operator remains responsible. A high score is not proof of a genuine fix.
• The interference map is MODELLED (pre-registered thresholds, not validated against a ground-truth measurement). A degraded cell does not identify interference as the cause; a cell not observed is not evidence of a clear route; it describes past reports and is not a forecast. It is aggregate only: no identifier is returned.
• Training NMEA is text only: no RF, IQ or waveform is synthesised, nothing is transmitted, and it is never for a vessel's live navigation systems.