Attestiq

Attestiq Validate · AI output assurance

Is the AI output correct? Prove it, every time.

Starting where errors cost real money and create examiner findings: card disputes. The same validator takes new policy packs for KYB, complaints, collections, and support agents.

Start an 8-week pilotRun Demo 1

What you get

Four surfaces, one evidence chain.

01

Validate API

POST an agent output plus its source records; get a verdict, a typed error list, citations, and signed evidence.

Used by · Engineering, on behalf of compliance

02

Evaluations

Run your agent's historical outputs through the validator: precision and recall by error type, top failure patterns.

Used by · Model risk, compliance

03

Review queue

Held outputs wait for a reviewer to release, block, or return them. Every label becomes training data.

Used by · Compliance analysts

04

Attestation pack

An examiner-formatted report: scope, agents covered, error rates, overrides, controls, integrity proofs.

Used by · Bank oversight, board

Typed errors

Every error names the rule it breaks.

A finding without a record reference or a citation is not a finding. Reviewers see the quote, the expected value, the observed value, and the evidence hash.

  • FACT.AMOUNT_MISMATCH

    A $68.40 purchase on a letter whose transaction record says $86.40.

  • FACT.REASON_CODE_MISMATCH

    A non-receipt dispute described under Visa fraud code 10.4.

  • DISCLOSURE.MISSING_REQUIRED

    A Reg E denial without the right to request the documents relied on (12 CFR 1005.11(d)(1)).

  • POLICY.TIMELINE_VIOLATION

    A Mastercard second presentment filed after the 45-day window.

  • POLICY.PROHIBITED_ACTION

    A retry policy that reattempts Visa category-1 declines.

  • PRIVACY.PAN_EXPOSED

    A full card number in a customer letter.

How a verdict is made

Code decides. Models read.

The validator works with no model at all. When a model is on, it extracts claims it must quote verbatim, judges issues it must cite to a pack rule, and writes a summary that may not introduce a number.

  1. 01

    Intake

    The original is hashed; personal data and card numbers are redacted before anything is stored or read by a model.

  2. 02

    Claims

    Amounts, dates, reason codes, card, merchant, and outcome — each with the exact quote it came from.

  3. 03

    Checks

    Fact, math, policy, and disclosure validators run against your records and the pinned policy pack.

  4. 04

    Verdict

    Code computes pass, warn, hold, or fail. Models may add cited findings; they never set the verdict.

  5. 05

    Evidence

    Every finding is a signed record written once, and every step lands on the tenant's audit chain.

For developers

Integrate in a day.

Call after each draft — synchronously, or asynchronously with signed webhooks. Or give your agent the validate_output tool over MCP.

  • REST API with idempotency keys and OpenAPI
  • Signed webhooks (HMAC, timestamped)
  • MCP server for Claude, OpenAI, and LangGraph agents
  • Python SDK with webhook verification
Open the demo console
curl -X POST https://api.attestiq.ai/v1/validate/validations \
  -H "X-API-Key: $ATTESTIQ_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "workload": "disputes.cardholder_letter",
    "agent":    {"id": "dispute-drafter", "version": "2.3.1"},
    "output":   "<the AI-drafted letter>",
    "records":  [{"type": "transaction", "id": "TXN-88213", "data": {...}},
                 {"type": "dispute",     "id": "DSP-2026-0918", "data": {...}}],
    "idempotency_key": "DSP-2026-0918-draft-3"
  }'

Eight weeks to an attested AI workload.

A fixed-fee pilot on one workload: an eval report on your historical outputs and an attestation draft for your compliance team.

Start a pilotExplore the demo console