Model Disclaimer

The wreckage-location engine produces a probability distribution over position, conditioned entirely on the evidence you supplied. A 90% containment contour means that, if the model's assumptions hold and your inputs are accurate, the wreckage lies inside the contour in nine of ten equivalent scenarios. It does not mean the wreckage is there.

Hypothesis rankings express relative consistency between a proposed mechanism and the evidence in the ledger. They are not probabilities of causation and carry no evidential weight of their own.

• Garbage in — an incorrect last known position, an inverted wind vector or a mis-timestamped fix will move the answer by tens of kilometres with no warning sign in the output.
• Unmodelled control input — a deliberate or sustained manual input after the last fix violates the sampled trajectory family and can place the aircraft entirely outside the contour.
• Drift over long durations — surface drift error grows superlinearly with elapsed time; multi-day searches should re-run with updated current fields rather than relying on the original field.
• Date-line debris backcasts — longitude is handled continuously, but the underlying drift remains two-dimensional; any crossing is labelled indicative and should be checked against specialist oceanographic modelling.
• Bathymetry resolution — GEBCO grid spacing is coarse relative to sonar swath planning; local terrain can invalidate a sector plan.
• Detection model — assumed probability of detection for a sweep is an estimate; an optimistic value causes the Bayesian update to wrongly downweight a swept area that was in fact searched badly.

Known failure modes and back-test results are documented in full on the Methodology page, including cases where the model performed poorly. We publish the misses deliberately.

Transcription of cockpit, ATC and interview audio and extraction from documents are machine-generated. They mishear, they hallucinate plausible words in noise, they lose speaker attribution under crosstalk, and they degrade on non-native accents and non-English speech.

Every AI output enters the ledger unverified and labelled. Verification by a competent human against the source recording is mandatory before the record is treated as evidence. Do not quote a machine transcript in a report without that verification.

Deployment of search vessels and assets, allocation of search effort, calling off a search, grounding an aircraft or fleet, issuing an airworthiness directive or safety recommendation, disciplinary action, prosecution, and settlement of a claim are decisions for the competent human authority. Platform output is one input among many and must never be the sole basis for any of them.

Search operations at sea carry risk to life. Treat the contour as a prioritisation aid subject to the judgement of the search coordinator, who holds information the model does not.

The physics and occurrence families cover, with tested worked examples: loss of control and stall, single and twin engine failure, propeller and turboprop behaviour including unfeathered drag and Vmca, icing, midair collision, controlled flight into terrain, fire, decompression, structural break-up and lifting-surface separation, high-energy dive, ditching, takeoff performance and runway excursion, fuel exhaustion and contamination, hypoxia and slow pressurisation loss, crew incapacitation, loading and centre-of-gravity errors, ground and ramp occurrences, turbulence injury without airframe damage, navigation interference, deliberate diversion, post-crash fire survivability on land, and wake turbulence.

Coverage is materially weaker for seaplanes and amphibians, tiltrotors, large remotely piloted aircraft, gliders, balloons, airships, ultralights, and military or experimental aircraft, and for cyber intrusion generally. Where a domain model exists, it refuses to produce a verdict when its governing inputs are absent and instead lists the missing evidence; treat a refused verdict as a direction to collect evidence, not as an inconclusive finding.

Survivability outputs are estimates conditioned on the impact and egress data entered. Where actual outcomes are known, treat the platform estimate as a planning aid, not a substitute for the medical and rescue record.

Deliberate-interference results test whether recorded behaviour is consistent with that hypothesis. They do not identify an actor, infer motive, establish criminal conduct or replace security and law-enforcement investigation.

The platform cannot interview investigators or witnesses, inspect wreckage, read out recorders, or independently establish any fact. Every fact it reasons over was stated by a person, and the reasoning is only as sound as those statements.

It does not and cannot determine probable cause. Probable cause is the formal finding of the investigating authority. The platform organises evidence, tests hypotheses for consistency, and surfaces gaps; the finding, the report and the safety recommendations belong to the authority and its investigators.

The engine has been validated to industry-standard rigour against a broad corpus of historical hull losses with known recovery positions — spanning deep water, shallow water, land, controlled flight into terrain, in-flight break-up and drifting ditchings. Every case is re-run using only the information searchers held at the time, with the recovery position withheld from the model and used solely for scoring.

The full results — containment rate, mean error, calibration curve, per-category error and baseline ablation — are published on the Validation page, refreshed as the model changes, and reproducible from the published proof pack. The dataset summary, methodology and step-by-step reproduction instructions are published on the validation evidence page at /validation-evidence. Validation performance is nonetheless retrospective and empirical; it does not guarantee performance on your occurrence, which may differ from the validated corpus in ways the evidence does not reveal.

To the maximum extent permitted by law, no warranty is given that any reconstruction is correct, that any hypothesis reflects actual causation, or that wreckage will be found within any contour. Liability is limited as set out in the Terms of Service.