Mandate Labs
model ward-1 · beta

Ward

A decision model for fraud, AML, compliance and risk. It takes a case and returns APPROVE, STEP_UP or DECLINE with a probability for each risk line, the signals behind it and what it could not check. It does not generate text.

ward-1model id. Pin a release or follow the alias
32K / 1Mtoken cap per real-time call / per file
11decision types behind one endpoint
3risk lines per answer: fraud, AML, sanctions
[01] The answer

One envelope. Three risk lines. Nothing hidden.

Fraud, AML and sanctions are answered separately, so a sanctions stop is never mistaken for a fraud score. Each line states its probability, its interval and how it was calibrated.

response · type: payment
{
  "model": "ward-1",
  "decision": "DECLINE",
  "outcome": { "choice": { "DECLINE": 0.58, "STEP_UP": 0.42 },
               "confidence": 0.37 },
  "risks": {
    "fraud": { "p": 0.598, "interval": [0.40, 0.87],
               "calibration": "reference_calibrated" },
    "aml": { "level": "escalate", "gates": false },
    "sanctions": { "status": "clear" }
  },
  "initiator": "cardholder",
  "signals": [ … ],
  "needs": [ … ],
  "not_assessed": [ … ]
}
  • decision. The field your code branches on. A decline from an earlier release is never loosened by a later layer.
  • outcome. The distribution over the options and a confidence, so a close call is marked as one.
  • risks. One line each for fraud, AML and sanctions, with probability, interval and the drivers.
  • initiator. Who started the payment: cardholder, merchant or agent. Declared or inferred. Conflicts fail closed to a step-up.
  • needs and not_assessed. The data that would sharpen the answer, and the checks that did not run.
  • Versions. Every answer names the model and the version of each component that produced it.

Field reference in the docs.

[02] Calibration

A probability should mean what it says.

Raw scores are mapped to probabilities with a reference map and shifted to your base rate. A line is labelled reference_calibrated only when its held-out calibration error is 0.05 or less.

Reference mapBlocksHeld-out ECEAUCStatus
Card fraud100.01040.877REFERENCE CALIBRATED
Statement AML140.03920.897REFERENCE CALIBRATED

Lines are refitted on each client's own labelled outcomes as they arrive. Set your base rate per client.

[03] Benchmarks

Measured on card authorization.

Two ISO 8583 suites: one where the cardholder has 25 to 50 past transactions, and a cold set with a profile and no usable history. Fraud stopped is measured at a 5% false-decline budget.

SuiteScorerAUCFraud stopped at 5% false declinesGenuine approved at 0.5Tokens / call
With historyWard0.98389 of 101298 of 298113
Cold startWard, with profile0.95255 of 65228 of 229238
  • Fewer false declines. Ward approved every genuine cardholder with history and 228 of 229 new ones.
  • Strongest on impossible travel, account takeover and first-time purchases.
  • Server time. 40 ms p50 and 55 ms p95 for the deployed model. Scoring itself is a fraction of a millisecond.

ISO 8583 card authorization benchmarks, October 2026.

[04] How it is built

Analyzers with published weights, then models fitted on your outcomes.

Ward is not one opaque network. Each decision type has its own analyzer whose signals and weights are documented. Learned layers are added on top only when they beat the analyzer on later cases.

analyzers

Typology analyzers

Card, account-to-account, statement, credit, merchant, account takeover and document analyzers. Every signal has a code, a weight and a typology it points to.

lists

Lists and intelligence

OFAC, UN, UK and EU sanctions lists, PEP and enforcement records, adverse media, and WMID, Ward's merchant intelligence data.

agent engine

Agent engine

Seven decision gates and two system steps, the KYA Score and the six EDQS-B behavioral detectors, for payments started by AI agents.

calibration

Reference calibration

Isotonic maps with intervals and a base-rate shift turn raw scores into probabilities you can set thresholds on.

client model

Per-client models

Scorecards and boosted trees trained on your labelled outcomes. They run in shadow, show their reasons and cut points, and decide only after maker-checker approval.

readers

Readers

A learned column reader for files in any layout and seven languages, MRZ and document forensics, and ISO 8583 data elements for card exports.

[05] Releases

Versioned, and every answer says which one.

GET /api/v1/models lists the served release, its components and the change log. Pin a version for stability or follow ward-1.

Nov 2026 · 1.5Separate risk lines and reference calibrationFraud, AML and sanctions answered as separate lines with probability and interval; client questions on a decision; account context on payments; per-client base rates.
Oct to Nov 2026 · 1.4Decision models per clientFitted, explainable models in the outcome loop with reasons and cut points; card-program ledgers; new statement signals from published typologies.
Oct 2026 · 1.3Files in any layout, merchant intelligence, pricingLearned column reader, ISO 8583 card exports, WMID, cardholder history in the request, evidence pack, and per-token pricing with a 1M-token window.
Oct 2026 · 1.0 to 1.2One endpointPOST /api/v1/decide for payment, intent, agent, screen, document and statement; then account takeover, credit and merchant types, async jobs and the governance pack.

Full changelog →

Try it on a case you already know the answer to.