Observatory results: 30-day Baltic back-catalogue and sensitivity study

Version kshana-observatory/adsb/v1 over kshana-interference-map/adsb/v2 (parameters pre-registered in the base method, unchanged). MODELLED: there is no external oracle, no ground truth and no cause attribution. A degraded cell means a high share of the aircraft sampled there reported low navigation accuracy that day; it does not say why. A cell that is not observed, or absent, was not assessed, which is not the same as clear.

What was run#

• Source: adsb.lol readsb history (ODbL 1.0), GitHub release v<YYYY>.<MM>.<DD>-planes-readsb-prod-0 of adsblol/globe_history_2026, fetched by scripts/fetch_adsblol_day.sh one part at a time (about 3.4 GB a day, 2 parts). The publisher states no checksums, so the SHA-256 of every part is recorded in back-catalogue/fetch/<tag>.parts.json.
• Region and days: the box 53.5, 9.0, 66.0, 31.0 (south, west, north, east: the Baltic Sea and its coasts), 30 UTC days from 2026-09-15 to 2026-10-14, 0.5 degree cells, by scripts/observatory_backcatalogue.sh (one day at a time, raw files deleted after each day).
• Release rules: the newest day (2026-10-14) was more than 24 h past its end when produced; the command refuses otherwise. Regional runs compute the equipment baseline and the day's background from the box's reports only (method.scope).
• Files: the 120 product files (GeoJSON, PNG, incident JSON, manifest per day) are not in git. The hash of every file is in back-catalogue/MANIFEST.json, and a day can be rebuilt from its archive and compared.

Example of what a day reads: on 2026-09-15 the archive held 110,619,797 position rows in 79,826 trace files; 103.0 million were outside the box, 5.6 million were not ADS-B, 1.2 million were on the ground or below 5,000 ft, 0.6 million carried no accuracy field, and 212,426 were aggregated (manifest.json input counters).

Reproducibility#

Three days (2026-09-18, 2026-09-26, 2026-10-03) were fetched again from the archive and run a second time: the GeoJSON, the PNG and the incident file are byte-identical to the first run (SHA-256 equal), so the product is deterministic even though the identifier hash salt is random per run.

Back-catalogue summary#

30 days: 96 degraded cells in total, 28 of the 30 days with at least one, 70 incidents (clusters of touching degraded cells), no day with a day-wide effect. The median cell share of affected aircraft (the day's background) is 0 on 29 days and 0.006 on one, so the background guard (a cell must exceed the background by 0.15) never changed a call in this region. The count of degraded cells per day runs from 0 to 8 (mean 3.2). About a third of the published cells each day are not_observed (too few aircraft sampled): coverage is the limit of this product, not the sky.

What this does not say: the days with more degraded cells are not "worse" in any physical sense, a persistent affected area would be understated (aircraft that fly only inside it all day have no baseline outside it and are not counted), and nothing here says who or what caused low accuracy.

Sensitivity study (three days)#

kshana interference-map sensitivity replays one day's recorded reports under one-at-a-time changes of the registered parameters (two values each, plus a 1.0 degree grid). Days: 2026-09-18 (the most degraded cells, 8), 2026-09-26 (1 degraded cell) and 2026-10-03 (0). Full reports: sensitivity/<day>.json (23 variants each; 2.2, 1.7 and 1.7 million rows replayed). Degraded cells under each variant (registered value in brackets; the registered result is 8, 1 and 0 degraded cells):

Reading it, without choosing a winner:

• The calls are sensitive to the accuracy-code thresholds and the sampling rules, and the small counts move a lot. low_nic_max, degraded_share and min_reports_per_aircraft_in_cell each change the number of degraded cells by a factor of two or more on every day; relaxing low_nic_max to 6 turns 1 degraded cell into 12 on 09-26 (Jaccard with the registered set 0.08). On 09-18 raising degraded_share to 0.4 keeps 1 of the 8 (Jaccard 0.125). A cell call near the thresholds is a statement about the thresholds as much as about the data.
• min_aircraft_sampled mostly moves cells between not degraded and not observed: on 09-18 it leaves the degraded count at 8 while 21 to 56 cells change state.
• Three guards are inactive here: background_margin, min_baseline_good_reports and day_confound_background change nothing on any of the three days (the background share is 0), so this region and period say nothing about how well they work. They would matter where a regional effect raises the background.
• A coarser grid (1.0 degree) changes the picture (2, 3 and 1 degraded cells; the published cells fall to about a third, 226 against 694 on 09-18, because each cell is four times larger): cell size is a method choice, and the observatory fixes it at 0.5 degrees.
• A day with no registered degraded cell (10-03) has degraded cells under 9 of the 22 non-grid variants (up to 8 with low_nic_max 6): "no degraded cell" is also a threshold-dependent statement.

No variant is a validated alternative, and the registered rule is unchanged by this study. The study does not say which parameter values are right.