Claim evidence review for insurers

A fabricated photograph cannot know what the weather station recorded.

Vericlaim reads the photographs and the claimant’s account, checks both against file forensics and independent records, and reports every finding with the innocent explanation beside it.

An aid to investigation, not a decision. The adjuster decides.

Example claim — not real data
DEMO-0001 · Vehicle collision · 3 photos
3 things do not add up — 2 in the photographs, 1 in the account.
98
Referral recommended
01
This photograph was submitted before, on another claim
It could be nothing, if The claimant may have resubmitted an old photograph in error, or this may be a supplement to the earlier claim rather than a new loss.
02
A photograph was taken before the incident
It could be nothing, if A device with an unset clock can be days out. The incident date may also have been reported approximately.
Every finding, the same way

Three answers, and the third is not a disclaimer.

An investigator who reads only the first two walks into an interview believing something a five-minute answer can dissolve. So the innocent explanation is printed beside the reason, at the same size, on every finding.

01
What does not add up

One sentence, in plain words. No detector names, no log-odds.

02
Why it stands out

The evidence behind it, cited: the photograph, the words in the account, or the record it contradicts.

03
It could be nothing, if

The most plausible innocent explanation, so the question to ask is obvious before anyone is accused of anything.

How a claim is checked

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

The model produces findings, each tied to cited evidence. A published rubric turns findings into a number, so the score can be reproduced, calibrated, and explained to a regulator.

0no model
File forensics

Metadata, provenance, and every image this carrier has seen before.

1no model
Independent records

Weather, daylight and the vehicle record at the stated time and place.

2vision model
Visual analysis

One pass per question, every observation pinned to image coordinates.

3judgment model
Narrative adjudication

Where the account and the evidence cannot both be true.

4judgment model
Adversarial defence

Every finding is argued against before it counts.

5no model
Scoring

A published rubric turns findings into a number. The model never does.

The full method

The score

One number, on a scale you can read.

Each band carries an action. None of them is “deny”.

No indicators
0–9

No action. Process normally.

Routine
10–29

Note in file. No investigation warranted.

Adjuster review
30–64

Put the suggested questions to the claimant before settling.

Referral recommended
65–100

Recommend referral to the Special Investigation Unit.

Principles

What it will never do.

These are constraints built into the system, not settings. There is nowhere in it to put a claimant’s name.

  • 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

“We will not quote an AI-detection accuracy number.”

Detecting generated images from pixels alone does not hold up on a generator the detector was not built against. Our defence is the record: a generated image cannot know the weather, the light, or the vehicle on file.

The score is provisional until it has been checked against real claim outcomes, and every report says so.

Honest limits

Put a claim through it.