Changed presentation
Four fictional report families use different names and wording, with varied formatting of their shared identifiers.
THE INVESTIGATION WORKSPACE
Follow the evidence.
Different names. Familiar patterns. Find what connects them.
FICTIONAL DEMOA sample investigation. All offers and identifiers are invented.
Evidence map
Identifier connections across independent reports
Your investigation starts here.
Add a message or screenshot to trace its identifiers.
Report queue
Source material, kept in view.
ON-DEVICE INTELLIGENCE
Find the familiar, even when the words change.
Neural similarity surfaces narrative leads. It never creates confirmed graph connections.
MiniLM · Runs on your device · Model downloads on first use
THE EVIDENCE REGISTER
Keep the receipts.
Review, correct, and trace every item in your workspace.
| Report | Channel | Identifiers | Connection status | Added |
|---|
No reports yet. Add your first piece of evidence.
MEASURED, NOT ASSUMED
Put the connections to the test.
A reproducible synthetic benchmark with difficult negative examples.
This evaluation measures identifier grouping on fictional, labelled data. It does not measure real-world fraud detection or the neural model.
What this test challenges
Misleading overlap
Unrelated offers share public form hosts and administrative email addresses. Those overlaps must not create corroborated groups.
Evidence integrity
Separate automated tests check duplicate suppression, OCR review gates, source validation, imports, and connection retraction.
No trained classifier is evaluated here. The baseline links any raw identifier overlap; ScamTrace requires reviewed, source-grounded evidence and corroboration.
A CLEAR CHAIN OF EVIDENCE
Trust what you can inspect.
ScamTrace connects reports. It does not pronounce people guilty.
Preserve the source.
Paste a message or read a screenshot with Tesseract OCR. Screenshot identifiers stay out of confirmed connections until you review their readings.
Show the relationship.
A shared payment destination, or both a shared email and phone, supports a corroborated report group. Other overlaps remain visible supporting leads. Public form hosts are excluded.
Keep it correctable.
Correct or exclude any extracted field. Connections and groups are recalculated immediately. Reposted content is counted once, using a normalized SHA-256 fingerprint.
What the AI does
Tesseract reads image text. A quantized MiniLM model produces semantic embeddings in a browser worker; cosine similarity surfaces related narratives separately from the identifier graph. Identifier extraction and connection policies are transparent deterministic code.
What the result means
Reports are user-supplied and unverified. A shared account establishes a shared destination, not a shared operator or proven fraud. Similar language is a weak lead. No connections found means insufficient evidence, not a safe offer.
Your evidence, your device
Evidence is stored in this browser’s IndexedDB. Text and screenshots are not uploaded to a server. Model files and fonts download from external providers. Export a case to back it up or share it intentionally; exported files can contain sensitive identifiers.