Forensic intelligence is judged on what arrives in the case file, not on what the marketing says. The walkthrough that follows is a single claim, scored end-to-end. The claimant, the vehicle, the photographs, the workshop are all synthetic. The forensic methods, signal vocabulary, scoring, API response shape, and case file structure are exactly as they appear in production.
A motor CASCO claim. Four photographs, three documents, one written statement. Submitted in good order. Scored in four seconds.
A first-party motor CASCO claim is filed through the carrier's mobile app. The claimant reports a single-vehicle collision with a tree on a rural road. There are no third parties, no police report, no witnesses. That is a common pattern in the most-claimed segment of the European motor market, and a pattern that gives a fraudulent submitter the most room to construct evidence.
The package contains seven items. To the claims handler, the package looks complete. To Veritura, the package is the input.
Every submitted artefact runs through every engine in the pipeline. No claim is sampled out, no engine is skipped, no analysis mode is partial. The pass takes four seconds; the four panels below are what the engine returned on this specific claim.
Of the four photographs in the submission, three returned forensic signals. The dual-model ensemble agreed on two of them. On the third, the two analytical paths disagreed; the disagreement itself is logged as a signal for the handler to weigh.
| Photograph | What surfaced |
|---|---|
| IMG_0241.jpg front-quarter damage |
Generative-model artefacts in the metal deformation region. Spectral signatures consistent with generative AI. Capture timestamp inconsistent. |
| IMG_0242.jpg VIN plate close-up |
Localized re-compression around the VIN plate area. Region duplication detected: portions of the plate duplicated and pasted from elsewhere in the same image. |
| IMG_0243.jpg rear damage |
Generative-model signature on full image. Sensor noise fingerprint inconsistent with the camera model declared in metadata. |
| IMG_0244.jpg interior |
No forensic signal. Image returns clean across the ensemble. |
The engine returns the artefact image for each finding and attaches it to the case file: the heatmap, the match overlay, the localisation map. The handler sees not only that the engine flagged something, but where.
The plate region exhibits compression-error inconsistencies relative to surrounding bodywork, with a copy-move match identifying duplicated character regions within the same image. Both signals fire together. The plate has been digitally substituted.
Each document passed internal consistency checks on its own. Cross-document analysis surfaced one conflict and one anomaly.
The written statement places the incident on a Tuesday evening. The repair estimate is dated Monday morning of the same week, before the incident the claimant describes. The two documents are inconsistent on a basic, falsifiable fact.
"… happened Tuesday evening, around 19:30, on the road toward …"
Estimate prepared: Monday, 11 May 2026, 09:14
The repair estimate is issued under a workshop name that does not appear in available trade registries for the declared region. The estimate is internally consistent; the issuer is not verifiable.
"AutoTech Premium Service Sp. z o.o."
No matching entity found in declared region.
The VIN on the registration document was extracted, normalized, and validated. The registration VIN is valid and present in the national vehicle registry. The vehicle is correctly registered to the policyholder.
The VIN visible on the plate in IMG_0242.jpg, after the duplicated region was excluded, returned a different number. That number is also a valid registry entry, but for a different vehicle, registered to a different owner, in a different region.
The registration document is real. The vehicle in the photographs is not the registered vehicle. The second VIN is associated with a different policy held by a different carrier.
Every image is fingerprinted on entry and indexed in the carrier's intelligence pool. The pool surfaces re-use across policies, channels, and time within the carrier's own data. No personal data crosses any boundary, and no new sub-processor relationship is created.
IMG_0241.jpg matched a perceptual fingerprint submitted to the same carrier through a broker channel eleven weeks earlier, under a different policy held in a different name. The image had been re-encoded and lightly transformed since the prior submission; the fingerprint match survived both.
The prior claim was paid.
Veritura returns a single 0–100 risk score and the full weighted composition that produced it. The score is the summary; the composition is the evidence. Both are in the API response, both are rendered in the handler's existing interface, both travel with the claim.
On this submission, the score lands in the high-risk band and triggers SIU referral. The composition is itemized below.
| Signal | Engine | Contribution |
|---|---|---|
| Generative-model artefacts (2 of 4 images) | Image forensics | High |
| Region duplication on VIN plate | Image forensics | High |
| VIN mismatch: registration vs. photographic evidence | Asset identity | High |
| Cross-channel fingerprint match against a prior claim | Cross-channel intelligence | High |
| Date conflict: statement vs. repair estimate | Document intelligence | Medium |
| Workshop not verifiable in trade registry | Document intelligence | Low |
| Sensor fingerprint inconsistency | Image forensics | Low |
Exact signal weights are operational intelligence. They are calibrated per carrier, validated against per-carrier claim corpora, and disclosed under NDA during technical due diligence. The composition is the evidence; the weights are not the product.
Claim proceeds through standard workflow. No required action from the handler.
Handler reviews the composition and decides: approve, request further information, escalate.
Handler refers to the SIU. The case file is assembled and attached to the referral automatically.
The handler always makes the final decision. Veritura returns evidence, not a verdict.
A high-risk referral is not a notification. It is a complete forensic file, assembled at the moment the score crosses into the high-risk band, attached to the case in the SIU's system of record, and structured for use without further evidence preparation.
Four artefact classes travel with every referral.
Every visualization produced by the engine: heatmaps, match overlays, localisation maps, signature charts, sensor-noise residue plots. Each rendered as a standalone image with the source claim reference, the engine that produced it, and a timestamp. SIU-presentable as standalone exhibits.
The full record of every signal that fired and every signal that passed, with the engine that produced it, the file it was derived from, and the threshold it cleared. The signal log is the structured form of the composition; it is also the document an investigator works through line by line when building a case.
The full JSON returned to the carrier's claims platform at the moment of scoring. Includes the score, the composition, the engine outputs, the artefact references, the version of every model that ran, and the elapsed time per engine. This is the immutable record of the analysis.
Every action the engine took, every file it read, every model it called, every result it returned, sequenced and timestamped, stored under a retention schedule of no less than twelve months. The trail is the document the SIU presents when a referral becomes a prosecution.
The package shape is identical across every high-risk referral the engine produces. Case-ready. No manual evidence assembly. No second screen. No second tool.
Veritura is designed to the standards that high-risk AI systems are held to in the EU: explainability, auditability, and a defensible record. That those standards apply to fraud detection is our judgement, not the regulation's. The score is decomposed. The composition is itemized. The artefacts are retained. The audit trail is immutable.
What an SIU investigator opens, when a Veritura-flagged claim arrives, is the work already done.
A thirty-minute scoping call. A sample claim cohort. API access.
We bring the rest.