Feedback

Fraud detection is adversarial, and the documents you see are not the documents we tuned on. Feedback is how the engine learns your issuers, your templates, and your risk appetite. Anyone whose API key can read a check can send feedback on it. There is no separate permission.

Send feedback

curl https://api.sphinxhq.com/api/v1/document-checks/dc_8f2c.../feedback/ \
  -H "Authorization: Bearer $SPHINX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "rating": "disagree",
    "corrected_decision": "clear",
    "reason": "Con Edison reissues statements monthly; the revision history is expected."
  }'

Field	Required	Notes
`rating`	yes	`agree` or `disagree`.
`reason`	no	Free text, up to 5000 characters. This is the field that actually improves the engine. A `disagree` without one tells us something is wrong but not what.
`flag_code`	no	Makes the feedback about one flag rather than the overall verdict.
`corrected_decision`	no	What the verdict should have been: `clear`, `pending`, `suspicious`, or `fraudulent`. Verdict-level only.

Verdict versus flag feedback

Omit flag_code and you are talking about the bottom line. Supply one and you are naming a specific detector:

{ "rating": "disagree", "flag_code": "font_inconsistency", "reason": "This issuer mixes fonts in the header on every statement." }

Flag-level feedback is far more actionable, because it points at one detector on one document family rather than at an outcome. Prefer it when you can identify what misfired. You can also report a flag we should have raised but didn’t. Pass the code from the catalogue with rating: "disagree". False negatives are the hardest signal for us to find on our own.

Response

{
  "id": 4218,
  "object": "check_feedback",
  "created_at": "2026-03-11T18:04:22Z",
  "document_check": "dc_8f2c...",
  "kind": "verdict",
  "rating": "disagree",
  "flag_code": null,
  "corrected_decision": "clear",
  "reason": "Con Edison reissues statements monthly; the revision history is expected."
}

Notes

Feedback is append-only. Changing your mind means sending a second row, not editing the first. We keep the sequence because a check that gets reconsidered twice is itself worth knowing about.
Each row records the engine’s verdict at the time it was given, so a later re-run of the same document never rewrites what you were reacting to.
Sending feedback does not change the check’s decision. Use the dashboard override for that.

Get started

Concepts

Send feedback

Verdict versus flag feedback

Response

Notes

​Send feedback

​Verdict versus flag feedback

​Response

​Notes

Send feedback

Verdict versus flag feedback

Response

Notes