Evidence for the validation claim
Validation evidence
Section 5 of the Model Disclaimer states that the wreckage engine is validated against a broad corpus of historical hull losses with known recovery positions, that every case is re-run using only the information searchers held at the time with the recovery position withheld from the model, and that the results are reproducible from a published proof pack. This page is the evidence for each of those statements, and it carries the same limits: validation is retrospective and empirical, and does not guarantee performance on your occurrence.
185 scored runs across 5 independent seeds.
Share of runs where the recovery position fell inside the claimed 90% contour.
Mean 17.64 km · worst case 165.2 km.
Largest movement of the solution point across seeds — the answer is not seed luck.
1. Datasets summary
37 scored occurrences spanning 1985–2023, of which 23 ended over water (to a maximum charted depth of 4,900 m) and 14 over land. 2 carry a published real-world swept area, which is what the swept-area reduction figure is measured against. Inputs per case are the last known state, the meteorological and oceanographic layers for the occurrence window, and the elapsed drift interval. The confirmed recovery position is loaded for scoring only and is never visible to the engine.
| Category | Cases | Over water | Median depth | Max drift interval | Span |
|---|---|---|---|---|---|
| Shallow-water loss of control | 9 | 9 | 55 m | 0 h | 1989–2021 |
| Deep-water loss of control | 8 | 8 | 2,200 m | 0 h | 1987–2016 |
| Land impact | 8 | 0 | — | 0 h | 1997–2023 |
| In-flight break-up | 5 | 3 | 36 m | 0 h | 1985–2015 |
| Terrain-masked / CFIT | 4 | 0 | — | 0 h | 2010–2016 |
| Ditching / surface drift | 3 | 3 | 15 m | 18 h | 1996–2009 |
Excluded from scoring
- MH370 Boeing 777-200ER · 2014-03-08 — No confirmed main wreckage position. Scored as 'no truth': the engine produces a prior, but there is nothing to score it against. Any product claim about MH370-class geometry is unvalidated.
- Various Southern Ocean SAR Mixed · — — Long-drift open-ocean losses where only floating debris was recovered months later. Debris beaching points constrain origin only weakly and are excluded from scoring.
Occurrences without a confirmed recovery position cannot be scored, so they are kept out of every hit rate on this site. They stay in the corpus so the engine's behaviour on genuinely unresolved geometry remains visible and reviewable.
Full per-case detail — flight, airframe, last known state, met layers, recovery position and the published source for each — is rendered live on the validation record, which re-runs the engine in your own browser rather than serving stored numbers.
2. Methodology
- Scoring is blind: the recovery position is withheld from the engine and used solely to measure error and containment after the run completes.
- Only information available to searchers at the time is supplied — no hindsight tracks, no post-recovery corrections.
- Each case is run across five fixed seeds so seed sensitivity is reported rather than hidden behind a single favourable run.
- Containment is scored against the contour the engine itself claimed, which is what makes the calibration curve meaningful.
- Baseline ablation compares the engine against a naive last-known-position circle, so the published gain is the gain over doing nothing clever.
- Cases where the engine performed poorly are published alongside the rest; the worst-case error above is a real case, not a trimmed outlier.
3. Reproducibility
The published pack is deterministic. Re-running it from source must produce the same canonical bytes and therefore the same SHA-256 digest; if it does not, either the engine or the corpus changed and the claim on this page is stale.
- Pack version
- 2.0.0
- Published
- 2026-08-27
- Canonical bytes
- 53,122
- Seeds
- 20260601, 7, 1337, 99991, 424242
- 01Fetch the proof
GET /api/public/validation-proof — no account, no key. The endpoint re-runs the whole pack from source on request and reports whether it reproduced the published digest byte-for-byte.
- 02Compare digests
Check the returned digest against 1ecd8b547553f637… and the canonical byte length against 53,122. Any divergent case is named individually in the response.
- 03Re-derive per case
Append ?pack=1 to receive the full pack: per-case, per-seed solution point, error, claimed radii, containment, sector rank and swept area — enough to recompute every headline number yourself.
- 04Re-run in your own browser
The validation record executes the same Monte Carlo engine client-side against the same inputs, so you can watch the numbers form rather than trust a stored table.
4. What this evidence does not establish
- It does not establish that wreckage will be found inside any contour on your occurrence. Containment is a property of the corpus, measured retrospectively.
- It does not validate geometry outside the corpus — long-drift open-ocean losses with no confirmed recovery position remain unvalidated, and are excluded from every figure above.
- It does not validate the inputs you supply. An incorrect last known position or an inverted wind vector moves the answer by tens of kilometres with no warning in the output.
- It says nothing about hypothesis ranking or AI transcription accuracy. Those are separate features, are not scored here, and remain unverified until a competent human verifies them.
- It does not transfer any decision to the platform. Search deployment, grounding, airworthiness action, recommendations and settlement remain with the competent human authority.
This page restates the Model Disclaimer and does not vary it. Where the two differ, the Model Disclaimer and the Terms of Service prevail.