M2Square

Feedback Loop

Submit analyst labels and use them as supervised learning input for future Sentra models.

Sentra improves through feedback. Analyst outcomes and partner labels are stored as feedback_labels and converted into a fraud target during training.

Submit feedback

curl "$SENTRA_API_BASE/v1/risk/feedback" \
  -H "Authorization: Bearer $SENTRA_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "tx_id": "m2txn_260605_001",
    "label_type": "fraud",
    "label_value": 1,
    "source": "manual",
    "note": "Analyst confirmed account takeover pattern.",
    "labeled_at": "2026-06-05T12:30:00Z"
  }'

Required role: Analyst or Admin.

Accepted labels

Label typeMeaningTraining target
fraudConfirmed fraud outcome.Positive when label_value=1, negative when 0.
chargebackChargeback-confirmed loss event.Positive when 1, negative when 0.
blockedConfirmed block-worthy event.Positive when 1, negative when 0.
false_positiveAnalyst confirmed the transaction was safe.Negative when 1; 0 is stored but ignored.
refundOperational refund outcome.Stored, ignored for fraud target.
otherCatch-all operational note.Stored, ignored for fraud target.

Accepted sources:

  • manual
  • system
  • partner

Training target resolution

During training:

  • Any positive target wins.
  • If no positive target exists but a negative target exists, the transaction is negative.
  • Ignored labels remain available for history but do not affect the fraud model target.

Console workflow

  1. Analyst opens /cases/{case_id}.
  2. Analyst reviews transaction context, score, reasons, evidence, AI status, and history.
  3. Analyst updates notes, status, and final decision.
  4. Analyst submits feedback label.
  5. Feedback appears in case history.
  6. Future model training reads feedback labels from the dataset or DB.

What feedback does not do

POST /v1/risk/feedback does not trigger training immediately. Training is a controlled Admin/model-ops action so the team can review data quality, label coverage, metrics, and model provenance before activation.

Before production learning, define:

  • What counts as confirmed fraud.
  • What counts as false positive.
  • Whether chargebacks are always fraud-positive.
  • Which system outcomes are trusted enough for source=system.
  • Whether compliance decisions should be blocked, fraud, or other.
  • How long to wait before converting open reviews into training labels.

On this page