Evidence · Walkthrough

What arrives at the SIU
when Veritura flags a claim.

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.

Synthetic walkthrough · No real claimants · No real vehicles · No real images
Step I · What the claimant filed

The submission, as it arrived.

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.

Claim reference CSC-2026-0418-XK
Declared incident date 12 May 2026
Declared loss € 8,420.00
Submission channel Mobile app · First-party

The package contains seven items. To the claims handler, the package looks complete. To Veritura, the package is the input.

PHOTO · 01
IMG_0241.jpg 2.4 MB · 4032 × 3024 · 12 May 2026, 19:42
PHOTO · 02
IMG_0242.jpg 1.8 MB · 4032 × 3024 · 12 May 2026, 19:43
PHOTO · 03
IMG_0243.jpg 2.1 MB · 4032 × 3024 · 12 May 2026, 19:44
PHOTO · 04
IMG_0244.jpg 2.6 MB · 4032 × 3024 · 12 May 2026, 19:45
DOC · REPAIR ESTIMATE
repair_estimate_0418.pdf 1 page · 184 KB
DOC · REGISTRATION
registration_scan.pdf 2 pages · 312 KB
DOC · STATEMENT
statement.txt 1 page · 4 KB
Step II · What the engine saw

The forensic pass.

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.

2.1
Image forensic ensemble

Image forensics

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.

Artefact 03 of 07 Returned · attached to case file
ela_imgo242_v1.png
Engine
Error Level Analysis (ELA)
Source
IMG_0242.jpg · 1920 × 1440 · 3.8 MB
Region of interest
x: 312 – 472, y: 220 – 270 (VIN plate)
Output
PNG · 1920 × 1440 · 2.4 MB
Co-firing engine
Copy-move forgery detection

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.

A single entry from the case file's artefact index. Every finding the engine returns is logged this way: filename, source, region, co-firing engines, confidence band, timestamp. The artefact image itself is attached to the case file.
2.2
Internal + cross-document analysis

Document intelligence

Each document passed internal consistency checks on its own. Cross-document analysis surfaced one conflict and one anomaly.

Cross-document conflict Document intelligence
Date conflict

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.

statement.txt · ¶3
"… happened Tuesday evening, around 19:30, on the road toward …"
repair_estimate_0418.pdf · header
Estimate prepared: Monday, 11 May 2026, 09:14
Issuer anomaly Document intelligence
Workshop not verifiable

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.

repair_estimate_0418.pdf · letterhead
"AutoTech Premium Service Sp. z o.o."
Trade registry lookup
No matching entity found in declared region.
2.3
VIN · Registration · Cadastre

Asset and identity verification

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.

FROM REGISTRATION DOCUMENT
WAUZZZ8E26A123456
Valid in national registry
Registered to: Policyholder of record
Region: Match declared
MISMATCH
EXTRACTED FROM IMG_0242.jpg
WBA3A5C50DF879214
Valid in national registry
Registered to: Different owner
Region: 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.

2.4
Carrier-scoped hash pool

Cross-channel intelligence

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.

−11 WEEKS −9 −6 −3 −1 TODAY PRIOR SUBMISSION IMG_0241.jpg FINGERPRINT MATCH Broker channel Different policy · Different name Mobile app channel Current claim

The prior claim was paid.

Step III · The composition

The score, and what it means.

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.

Low Standard workflow
Medium Handler review
High SIU referral
THIS CLAIM HIGH RISK · SIU REFERRAL
Signal composition on this claim
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.
Risk bands, recap
Low

Claim proceeds through standard workflow. No required action from the handler.

Medium

Handler reviews the composition and decides: approve, request further information, escalate.

High

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.

Step IV · What the SIU receives

The case file, as it arrives.

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.

01

Forensic artefact images

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.

02

Signal log

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.

03

API response payload

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.

04

Chain-of-custody audit trail

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.

Explainability · Audit · Defensibility

Every signal that fired.
Every signal that passed.

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.

Open the conversation

A thirty-minute scoping call. A sample claim cohort. API access.
We bring the rest.

Contact