An evaluation is a stored verdict on a subject, and its ID starts with eval_. Intelligence is its own product: you can screen people, accounts and payouts without sending money through Anton, and accounts with no payout surface at all use exactly this call.
The credential needs the intelligence scope. Authentication covers minting the DPoP-bound token. The examples leave out the per-request DPoP proof header; send it with every call.
Say who you are screening
The entity is the subject, and entity.type is person or business. Give it something to search on: name or vault_token_id. Add country and, for a person, birth_date as YYYY-MM-DD. Details you leave out can come back as the reason information_incomplete.
reference_id is your own identifier for the subject. client_reference_id, on the request itself, is your identifier for this evaluation.
Create the evaluation
use_case tells Anton what the screening is for, and each value has its own required fields: payee_registration here, with instrument_screening, payout_screening and merchant_onboarding in their own guides. Send an Idempotency-Key; it is required on this endpoint, and a retry with the same key returns the first answer instead of running the screening again.
Every use case except payout_screening answers 201 with the evaluation already decided.
curl -X POST \ https://api.antonpayments.dev/v1/intelligence/evaluations \ -H "Authorization: DPoP $ANTON_ACCESS_TOKEN" \ -H "DPoP: $DPOP_PROOF" \ -H "Idempotency-Key: intel-payee-84012" \ -H "Content-Type: application/json" \ -d '{ "use_case": "payee_registration", "client_reference_id": "crm-payee-84012", "entity": { "type": "person", "reference_id": "payee-84012", "name": "Maria Testperson", "country": "DE", "birth_date": "1988-03-14" }, "metadata": { "crm_id": "42" } }'
metadata is yours: strings only, up to 50 keys. It is stored on the evaluation and returned with it, never searched and never used in screening.
Read the verdict
Branch on decision: approve, review or decline. score runs from 0 to 1000 and higher is worse; tier is its band. reasons is empty for a clean approval, and otherwise says why it isn't one. required_actions says what you can do about it. Decisions lists every value.
coverage tells you how much of the screening answered. Treat partial or error as a weaker verdict than complete, and don't approve on it silently.
The result fields exist only once status is completed. A failed evaluation has no verdict, only failure_reason: "evaluation_failed"; run it again with a new key.
Read it back
Evaluations belong to your account. GET /v1/intelligence/evaluations/{id} returns one, and GET /v1/intelligence/evaluations lists them newest first in the standard page shape. An ID that is someone else's answers 404, the same as one that doesn't exist.
Check a request without running it
POST /v1/intelligence/evaluations/validate takes the same body and answers valid: true with the event_type it maps to, or the same field errors as the real call. It stores nothing, so it needs no Idempotency-Key. It suits CI.
When it goes wrong
Idempotency-Key header.use_case is missing, or a value is invalid. error.details lists each one.intelligence scope.intelligence capability.Retry-After and resend.Rate limits and errors has the limits and the full list.
Next
People and businesses, with direct details or a vault token.
Bank accounts, IBANs, cards and crypto wallets.
The one use case that answers 202 and finishes later.
Details sent directly are used for the check and then discarded; references point at data already in the vault.