Risk evaluation

The single most important endpoint in Anti-Fraud Platform. Call it on every onboarding application and every transaction.

Request

Common data fields (transaction lane)

Common data fields (onboarding lane)

data is intentionally open-ended — rules read whatever fields they reference, so you can pass everything you have.

Response

{
  "data": {
    "decision": "allow | flag | review | block",
    "fraudScore": 0-100,
    "subScores": {
      "rules": 0-100,
      "velocity": 0-100,
      "ml": 0-100
    },
    "appliedRules": [
      {
        "code": "VL-003",
        "name": "High amount unverified counterparty",
        "action": "review",
        "severity": "high",
        "score": 20
      }
    ],
    "sessionId": "alt_01HXY..."
  },
  "ok": true
}

fraudScore is the headline figure — a weighted blend of subScores. Use it for downstream analytics; use decision for the actual gate.

Decision logic

       ┌──────────────┐
       │   subScores  │
       │  rules: R    │
       │  velocity: V │
       │  ml: M       │
       └──────┬───────┘
              ▼
       fraudScore = weighted(R, V, M)   ← weights configurable per-org
              │
              ▼
       applied rules' actions are merged.
       If a rule with action "block" fires AND its bypass-ML flag is true, decision = "block".
       Otherwise:
         fraudScore ≥ blockThreshold  → block
         fraudScore ≥ reviewThreshold → review
         fraudScore ≥ flagThreshold   → flag
         else                          → allow

Default thresholds: flag 30 · review 50 · block 75. Configurable via dashboard.

ML suppression

Rules can be marked bypassMl: true. These enforce their action regardless of ML confidence — useful for hard regulatory rules (blocked country, blocked NIK).

Non-bypass rules may have their action downgraded when ML confidence is below the org's per-lane mlThresholds:

block  → review  → flag  → allow

This prevents over-blocking when the ML model is uncertain. Disable by setting mlThresholds.onboarding = 0 (never downgrade) or mlThresholds.transaction = 0.

Idempotency

Pass externalId. Within 24 hours, replays return the original decision with data.idempotent: true.

Test mode

For unit-testing without persisting alerts:

Accepts the same body plus an optional scenario:

{
  "lane": "onboarding",
  "scenario": "clean | medium | high",
  "payload": { /* same shape as /api/evaluate */ }
}

scenario injects a fake risk score (40 / 60 / 90) to deterministically force a decision path. Useful for testing your downstream handling without depending on actual rules firing.

Latency expectations

Sandbox is configured for low fixed cost, not low latency — expect p95 ~120ms.

Production p95 SLAs:

Per-tier contracts are in your MSA — talk to sales@quantumelixir.tech if your latency budget is tight.