Fraud detection. A score, with the receipts.

Overview

A document-level risk score plus the evidence that produced it.

• 0-100 score — calibrated against the labelled samples in your tenant; higher is more suspicious.
• Per-signal verdicts — each fired signal carries a category, severity and a region-of-interest reference.
• Synthetic-generation checks — looks for diffusion artifacts and known generator fingerprints, not just copy-paste tampering.
• Audit-ready — every score links to a trace you can show an auditor or regulator.

Endpoint

Request

curl -X POST https://api.cogneris.ai/v1/fraud/score \
  -H "Authorization: Bearer $COGNERIS_KEY" \
  -F "file=@./claim-photo-front.jpg" \
  -F "context=insurance_claim" \
  -F "signals=all"

Parameters

• file or source.url — image, scan or PDF.
• context — hints the calibration: invoice, kyc, insurance_claim, payslip, generic.
• signals — all (default), or a comma-separated subset like tampering,metadata.
• linked_documents — array of prior document IDs to cross-check (e.g. invoice ↔ purchase order).

Response

{
  "data": {
    "score":    72,
    "verdict":  "review",
    "findings": [
      {
        "signal":    "image_tampering",
        "category":  "tampering",
        "severity":  "high",
        "page":      1,
        "bbox":      [412, 680, 724, 740],
        "detail":    "Pixel inconsistencies around the date field; JPEG quantisation mismatch."
      },
      {
        "signal":    "metadata_anomaly",
        "category":  "metadata",
        "severity":  "medium",
        "detail":    "Author tag references an editor not present in genuine samples."
      }
    ]
  },
  "meta": {
    "job_id":   "fr_01J9PR1Z…",
    "model":    "flx-fraud-2026-04",
    "audit_url": "https://app.cogneris.ai/audit/fr_01J9PR1Z"
  },
  "has_errors": false
}

Signal catalogue

Thresholds & routing

• 0 – 29 · pass — green-light the workflow.
• 30 – 69 · review — route to a human; show them the findings list.
• 70 – 100 · block — hold the document; require a second-level reviewer.

Bands are tenant-tunable. Override defaults via PATCH /v1/fraud/thresholds.

Feedback loop

When a human reviewer overturns a score, post the verdict back via POST /v1/fraud/feedback. Confirmed-fraud and confirmed-clean labels both feed the next calibration cycle, run weekly per tenant. We never share fraud labels across tenants.

Errors