Introduction

Send the input — a payment, a name, an ID document, a bank statement, an applicant — and get one decision your code can act on: APPROVE, STEP_UP or DECLINE, with the score, the signals that fired and the typologies they point to.

Ward is Mandate Labs' decision model for fraud, AML, compliance and risk. Fraud and compliance teams usually stitch together a rules engine, a screening vendor, a document vendor and a spreadsheet of thresholds, then write glue code to turn each answer into an action. Ward answers all of those questions through one endpoint and one answer shape, so your code branches on one field.

input ──► POST /api/v1/decide ──► { decision, score, signals, typologies, result }
                                        │
                     APPROVE → proceed   STEP_UP → ask / review   DECLINE → stop

Every answer says what it saw and what it could not see. A list that was not checked is reported as not_checked, never as clear; a component that could not run is a reason to review, never a pass.

Ward capabilities

Each capability is a decision type. Send it to POST /api/v1/decide with type, or let Ward infer the type from the input.

Type	Answers	Typical input
`payment`	Should this payment go through? Agent or person, card, A2A or crypto	amount, merchant, card or agent context
`intent`	Is this AI agent's intended action within its mandate?	agent, session, intent, reasoning
`agent`	Can this AI agent be trusted right now? (KYA)	`agent_id` or DID
`screen`	Is this person or company on a sanctions, PEP or enforcement list?	name, date of birth, country
`document`	Is this ID document consistent and unaltered?	MRZ lines and/or image or PDF
`statement`	What AML and fraud risk does this account history show?	bank statements, card exports, alert files
`ato`	Was the account taken over before this disputed transaction?	disputed transaction + account logs
`credit`	Can this customer carry the obligation, from their statements?	statements + requested instalment
`merchant`	Is this merchant safe to onboard or keep?	business, owners, observed rates
`kyc`	Is this applicant verified? Identity, email, phone, document	applicant JSON

Ward Ask answers follow-up questions about any result ("which counterparty matched a list?") from the result's own data.

One answer shape

{
  "id": "dec_01JC4Z...",
  "object": "decision",
  "model": "ward-1.4.6",
  "type": "screen",
  "decision": "STEP_UP",
  "score": 0.91,
  "score_kind": "name_similarity",
  "signals": ["SCREEN_POTENTIAL_MATCH"],
  "typologies": [],
  "needs": [],
  "result": { "...": "the capability's full answer" },
  "context": { "window_tokens": 1000000, "used_tokens": 412, "used_pct": 0.04 },
  "latency_ms": 38.2
}

See The decision envelope for every field.

Next steps

Quick startYour first decision in five minutes. How Ward decidesWhat a decision model is and what it is not. CapabilitiesOne page per decision type, with inputs and outputs. PatternsStep-up routing, fail-closed handling, review queues. API referenceEvery field of POST /decide. SDKsPython and JavaScript clients.

An index of these pages for coding agents is at /llms.txt.