Warddocs
Models

Models

The current release is ward-1.5.9 (November 13, 2026). GET /api/v1/models lists the releases, the version of every component and the changelog.

curl https://api.mandatelabs.ai/api/v1/models -H "X-API-Key: $WARD_API_KEY"

Choosing a model

Send "model" on /decide. ward-latest (the default) always serves the newest release; older ids are aliases of the release that superseded them. An unknown id answers 422 unknown_model with the list of valid ids.

Every answer says which release answered (model) and which component versions ran (components), for example:

Component Version
statement_matrix stm-matrix-7
statement_reader stm-reader-4
card_analyzer card-analyzer-1.3
screening screen-4
credit_risk credit-3
merchant_risk merchant-risk-2
ato_review ato-review-1
ask_router ask-2026-10-21-…
typology_catalog 1.0.0, card-0.3.0, crypto-0.2.0, financial-0.2.0

ward-2 (opt-in)

Send "model": "ward-2" to get the corrected decision mapping; ward-latest stays on ward-1 and its answers do not change. ward-2 is available where risk lines are enabled. It changes three answers, and every answer it changes carries decision_ward_1, what ward-1 would have decided:

Type ward-1 ward-2
statement, band high, no strong sanctions match DECLINE STEP_UP; the AML line says escalate
ato, takeover likely DECLINE (meaning: accept the customer's claim) STEP_UP; outcome is likely_account_takeover
payment with account.movements, AML priority escalate No effect At least STEP_UP (aml_gated: true), for accounts with the AML gate enabled. Never DECLINE from AML alone

result.assessment.fraud_probability is deprecated in both: read risks.fraud (raw_score, and p with its score_kind).

Recent changes

Release Date Highlights
ward-2.0 2026-11-12 Opt-in alias with the corrected decision mapping (above)
ward-1.5.9 2026-11-13 Cards: stolen card details used online from a new device to a new delivery address; a cardholder who passes 3-D Secure on a new phone is no longer read like a takeover (unless a SIM swap or a passcode to a new number says otherwise). Bank transfers: a call or remote session while paying someone already paid, at the usual amount, weighs less; a growing series to a recent payee weighs more
ward-1.5.8 2026-11-13 Bank transfers: scams paid in a series (romance, investment), a first payment to a crypto exchange and a dormant account paying out; fewer stops on confirmed business payees and supplier account changes; calibrated transfer probabilities with and without the payer's history
ward-1.5.7 2026-11-12 Generic business names are never a strong sanctions match on the name alone; the matched alias is named; questions agree with their risk line; three more initiator conflicts
ward-1.5.6 2026-11-12 Initiator conflicts: an agent identifier alone contradicts a human declaration; merchant on a present card is a conflict
ward-1.5.5 2026-11-12 questions on /decide (beta, when enabled); typology path on the fraud line
ward-1.5.4 2026-11-12 AML on a single payment from account.movements (advisory)
ward-1.5.3 2026-11-12 Decision as a Choice with a probability per option; confidence
ward-1.5.2 2026-11-12 Reference-calibrated probabilities with intervals; base rates per account
ward-1.5.1 2026-11-12 Bank transfers: sanctions and jurisdiction leave the fraud score; an unscreened payee steps up
ward-1.5 2026-11-12 Risk lines (risks, outcome, answers, not_assessed) behind a flag
ward-1.4.6 2026-11-02 Tie-safe cut points for client models
ward-1.4.5 2026-11-02 Statement matrix stm-matrix-7: four signals from published typologies (aggregate structuring, balance drain, escalating payments, ATM runs), observed only
ward-1.4.4 2026-11-01 Hard stops split by strength when a client model decides; potential sanctions matches become review floors
ward-1.4.3 2026-10-27 Card analyzer: new-merchant spree
ward-1.4.2 2026-10-27 Attempted vs settled amounts in card-program ledgers

The full history is in the changelog of GET /api/v1/models. The model card (intended use, data, validation, limitations, change control) is available on request.