Outcomes
Availability: turned on per API deployment; until then these endpoints answer 404. Ask Mandate to enable it for your account.
Every answer carries id and ref (wref_...). Keep either, or none: an outcome can also name the decision by your own reference or by the network's.
Name the decision
| Field on the outcome | Matches |
|---|---|
decision_id |
Ward's decision id |
ref |
The ref Ward returned |
reference |
Your id sent as reference on /decide, or transaction_id, client_reference, external_ref, case_id in the input |
card |
{"rrn", "stan", "auth_code"} together |
network_reference |
The card network reference |
crypto |
{"chain", "tx_hash"} |
Identifiers are kept as keyed hashes, never in clear. One decision: linked. None yet: kept and linked when the decision arrives. Several: 409 ambiguous_reference with the candidates; Ward never guesses.
Send outcomes
curl https://api.mandatelabs.ai/api/v1/ward/outcomes \
-H "X-API-Key: $WARD_API_KEY" -H "Content-Type: application/json" \
-d '{"card": {"rrn": "612345678901", "stan": "000777", "auth_code": "AB12CD"},
"event_type": "chargeback", "network": "visa", "reason_code": "10.4",
"occurred_at": "2026-11-02T10:00:00Z", "source_event_id": "cb-88231"}'
Bulk: POST /api/v1/ward/outcomes/bulk with {"items": [...]}, up to 5,000. The same event (same source_event_id, else the same references, type, day and code) is stored once.
event_type: chargeback, chargeback_won_by_merchant, chargeback_lost_by_merchant, fraud_report, scam_report, customer_cancelled, account_closed (with reason: mule, money_laundering, ...), sar_filed, suspected_flag, cleared_by_review, refund, other. client_label takes your own word, mapped by your label dictionary.
The label Ward derives
GET /api/v1/ward/decisions/{id}/outcomes shows the events and the label (rule labels-1):
| Events | Label | Trains |
|---|---|---|
| Fraud chargeback | fraud, provisional | after 45 days with no merchant win |
| ... then the merchant wins | first-party misuse | yes |
| Consumer dispute, processing or authorisation chargeback | dispute | no (not fraud) |
| Fraud or scam report, SAR, account closed for mule activity | that label | yes |
| Suspected flag alone | suspected | never as confirmed |
| Cleared by review | genuine | yes |
| No outcome, an approved card payment over 120 days old | genuine (matured) | yes |
| No outcome, an approved transfer over 13 months old (your policy) | genuine (matured) | yes |
AML monitoring never becomes genuine by silence. Set your own windows with PUT /api/v1/ward/thresholds: {"maturity": {"card_days": 180, "transfer_days": 365}} (card from 120 to 540 days, transfers 30 to 800).
Decisions and the queue
GET /api/v1/ward/decisions: your decisions, newest first, filtered bytype,decision,label_state,purpose,since,until; page withbefore.GET /api/v1/ward/outcome-queue: stepped-up decisions and the review sample of approvals with no outcome yet, outcomes not matched, words your dictionary does not map, and suspected replays.POST /api/v1/ward/decisions/{id}/replay-review{"clear": true}: a production call whose event happened long before Ward saw it, or whose reference Ward saw on an older decision, is held out of your model's training until you clear it.
Learning settings
GET / PUT /api/v1/learning/settings: train_on_outcomes (your own model trains on your outcomes; on by default) and review_sample (2% of approvals go to your queue; with it off your models are marked as trained on stopped cases only).