How it works

Six tiers. Three of them never call a model at all.

The model never produces the score.

It produces discrete findings, each tied to cited evidence. A deterministic rubric turns those into a number. A score that comes out of a model cannot be calibrated, cannot be reproduced, and cannot be explained to a regulator. A score computed from findings by a published rubric can be all three.

The tiers

0
File forensicsno model
Metadata, cryptographic provenance, perceptual hashing against every image this carrier has seen before, compression history. Runs at full resolution. A vision model cannot read an EXIF timestamp or verify a signature — these are the checks that hold up when challenged.
1
Independent recordsno model
Weather at the loss location and time, solar geometry computed from the timestamp, vehicle identification. A fabricated photograph is unmoored from the world: it cannot know what the weather station recorded. This is the layer that does not decay as image generators improve.
2
Visual analysisvision model
One pass per question rather than one pass per image. Damage mapping, age of damage from full-resolution crops, scene and lighting, internal coherence. Every observation carries image coordinates so it can be verified.
3
Narrative adjudicationjudgment model
Where the account and the evidence cannot both be true. Each contradiction cites the exact words and the evidence that falsifies them. Nothing is inferred from how the account is written — length, fluency and detail track education and trauma, not honesty.
4
Adversarial defencejudgment model
Every finding is argued against before it counts. A finding a competent person can explain away is not evidence, and acting on it costs a real claimant a real investigation.
5
Scoringno model
A published rubric converts findings into a number: per-category caps so correlated detectors cannot stack, a discount for uncorroborated findings, and hard guards that prevent a referral on inference alone.

What we do not do

  • NeverState that a claim is fraudulent
  • NeverRecommend denial of a claim
  • NeverContact a claimant
  • NeverReceive a name, address, date of birth or any personal identifier
  • NeverReason from a claimant’s language, education or manner of writing
  • NeverRetain a raw photograph beyond 24 hours

Honest limits

The scoring model is not yet calibrated. Calibration requires claims with confirmed outcomes, which requires a carrier. Until then every report carries a banner saying so, and no accuracy figure should be quoted from this system as a probability of fraud.

Detecting AI-generated images from pixels alone does not work reliably. Every such detector collapses on a generator it was not built against, and published accuracy figures are always measured in-distribution. Our defence is tier 1: a generated image fails against independent records. We will not quote an AI-detection accuracy number.

Three detectors are switched off. They were measured against 2,672 licence-verified real photographs and found to have no discriminative power. One of them ran backwards. They remain documented rather than deleted, so the test and its result stay visible.