Frequently asked questions
Can a model pass or fail a check?
No. Only the deterministic, versioned rules engine emits PASS, FAIL, or CANNOT_VERIFY. Model output can add context to an investigation or draft an explanation, but it cannot create, upgrade, or dismiss a verdict.
What happens when data is missing or disputed?
The affected checks resolve to CANNOT_VERIFY — never an assumed pass. Runs waiting on required inputs hold in waiting_for_input, missing sources alert, and no materiality threshold clears a CANNOT_VERIFY.
Does Definite write to my systems or file for me?
No. Integrations are read-only, Definite never initiates transactions, and it never files with a regulator or counterparty. Draft exports are watermarked DRAFT - NOT FOR FILING; final exports require a named admin or owner and are downloaded and submitted by your team.
Does the agent runtime store my documents?
No. It is stateless — no database, no stored source documents. It receives only the declared fields a task needs, for the duration of the request. Durable state (runs, receipts, snapshots, artifacts) lives in your tenant's isolated Vault storage.
Is my data used to train models?
No. Client data is never used for model training. Agent tasks run against an approved per-tenant model catalog through cloud provider endpoints, and model involvement is disclosed in run telemetry.
Which models does Definite use?
New tenants default to DeepSeek V3.2 via AWS Bedrock, with MiniMax M2.5 and Claude Haiku 4.5 in the approved catalog. Tenant admins choose from the catalog; models never emit verdicts regardless of which is selected.
What exactly is in a receipt?
Rule ID, version, and citation; the verdict; typed input references (opaque IDs plus sha256: content hashes); the previous receipt hash; the signed payload, signature, algorithm, and signer key. Raw sensitive values are never present. See independent verification.
Can I verify evidence after leaving Definite?
Yes. Receipts and proof bundles verify with the published JWKS keys using standard hashing and signature checks — no Definite availability required. Checkpoints are externally witnessed, so ledger tampering is detectable from outside.
Who can approve things?
Named humans with the right role. Reviews, mapping activations, definition publishing, and final exports require authenticated admin or owner actors through the console; a raw API key cannot perform them. Reviews never rewrite an existing verdict — corrected data produces a new run with new receipts.
What happened to the older policy-evaluation APIs?
Endpoints from earlier product generations return 410 GONE. The current surface is readiness, programs, runs, exceptions, receipts, ledger proofs, and the export destinations documented here.