Evidence, not guesswork
Every ThisVerify risk score is built from findings you can point to: a field in the file, a region on the document, or a calculation that does not add up.
1. Intake and fingerprint
File type detected from content (not the extension), SHA-256 fingerprint, encryption and private storage.
2. Metadata
Producer, Creator, XMP, EXIF, creation and modification dates, colour profile and image origin.
3. File structure
Save cycles, annotations, form fields, embedded fonts and font subsets, digital signatures and ByteRange, white boxes over text.
4. Pixels
Error level analysis on 16×16 blocks, statistically anomalous regions grouped into areas.
5. Visual analysis
Vision models — Google Gemini and, optionally, your private open model — inspect the document at high resolution against a checklist and a knowledge base of issuers, and extract the data.
6. Logic and history
Mathematical validation of the extracted data, an optional reasoning auditor, and a check whether the same file was submitted before.
How the risk score is computed
Every finding has a severity (info, low, medium, high, critical) and a confidence. Findings are combined as independent probabilities: each one removes part of the remaining trust in the document.
- Low risk — no meaningful signs found.
- Needs review — signs that justify a closer look.
- High risk — clear evidence of editing, contradiction or fabrication.
The report shows exactly how much each finding contributed to the score.
What the system does not do
We believe in transparency. ThisVerify supports decisions; it does not replace confirmation with the issuer:
- It does not connect to banks and cannot confirm that money was actually transferred.
- Photos of photos, heavy compression or low resolution reduce the ability to detect edits.
- A sophisticated fake built from scratch on an exact template may pass visual review — which is why the logic and history layers matter.
- The score is not a legal opinion or admissible forensic evidence.