Back-test / hull-loss localisation
Every case below is re-run through the same Monte Carlo engine the platform uses live, seeded and deterministic. The model receives only what searchers held at the time — last known flight state, ambient winds, surface current, water depth. The confirmed recovery position is withheld from the model and used solely to score the result.
| Flight | Date | Error | R50 | R90 | Containment | Sector | Verdict |
|---|---|---|---|---|---|---|---|
AF447 Airbus A330-203 | 2009-06-01 | 31.6 km | 32.2 km | 57.1 km | 50% | — | hit · r50 |
QZ8501 Airbus A320-216 | 2014-12-28 | 39.4 km | 26.0 km | 43.9 km | 86% | — | hit · r90 |
JT610 Boeing 737-8 MAX | 2018-10-29 | 0.6 km | 4.4 km | 17.8 km | 0% | S2 | hit · r50 |
MS804 Airbus A320-232 | 2016-05-19 | 34.6 km | 32.3 km | 56.3 km | 57% | — | hit · r90 |
KI574 Boeing 737-4Q8 | 2007-01-01 | 20.0 km | 29.2 km | 50.8 km | 26% | — | hit · r50 |
SJ182 Boeing 737-524 | 2021-01-09 | 15.9 km | 10.1 km | 21.8 km | 79% | — | hit · r90 |
MI185 Boeing 737-36N | 1997-12-19 | 19.6 km | 29.0 km | 50.3 km | 26% | — | hit · r50 |
GE235 ATR 72-600 | 2015-02-04 | 0.7 km | 1.5 km | 5.2 km | 21% | S1 | hit · r50 |
IY626 Airbus A310-324 | 2009-06-30 | 4.2 km | 2.9 km | 10.2 km | 68% | — | hit · r90 |
Deep-ocean loss of control from cruise after unreliable airspeed. The real search burned two years across ~17,000 km²; the debris field sat ~6.5 NM from the last ACARS fix.
Main debris field, located April 2011 by Remus AUVs after three failed campaigns. Source: BEA final report f-cp090601 (last ACARS position and wreckage site).
Upset and stall in convective weather over shallow water. Recovery site lay ~28 km south of the last radar return.
Fuselage located on the Java Sea floor, January 2015. Source: KNKT final report KNKT.14.12.29.04.
High-energy near-vertical impact from low altitude — the tightest possible prior. A good engine must not over-disperse this one.
Main impact point in the Java Sea, ~34 m depth. Source: KNKT final report KNKT.18.10.35.04.
Cruise fire event with a long, poorly-constrained final descent — the hardest of the set and the one that stresses the heading-spread model.
Recorder and main wreckage field, eastern Mediterranean, ~3,000 m depth. Source: Egyptian AAIC investigation and John Lethbridge survey positions.
Spiral dive from cruise into deep water. Recorders were pinpointed almost directly beneath the last radar plot.
Recorders located on the Makassar Strait floor, ~2,000 m depth, August 2007. Source: KNKT final report KNKT/07.01/08.01.36.
Autothrottle asymmetry roll upset shortly after departure. Short, steep and shallow-water — the case where a tight prior directly saves search days.
Main impact site off Pulau Laki, Java Sea, ~23 m depth. Source: KNKT preliminary/final report KNKT.21.01.01.04.
Near-supersonic dive from cruise onto land/river — the land-surface control case, where no drift term applies and dispersion must stay disciplined.
Impact into the Musi River near Palembang; shallow riverbed recovery. Source: NTSC Indonesia investigation and NTSB comments (1997).
Very low altitude, low energy urban loss of control. The engine must not blow this into a multi-kilometre zone.
Impact into the Keelung River, Taipei, after engine shutdown and stall. Source: ASC Taiwan final report ASC-AOR-16-06-001.
Night go-around upset near the coast with an 18-hour surface drift window — the case that exercises leeway and current advection rather than descent physics.
Debris field north-west of Grande Comore; recorders recovered from ~1,200 m. Source: Comoros / BEA-assisted investigation, published search positions.
Re-runs every benchmark across a fixed seed sweep and emits a canonical, hash-bound JSON artefact: full inputs, engine configuration, per-seed solutions, error, containment and search area, plus seed-sensitivity of each solution point. The SHA-256 digest is computed over the canonical body with the generation timestamp excluded, so an auditor who re-runs the same pack version must reproduce the identical digest byte-for-byte.