AI can be right — and still trigger the wrong action.

Zorelan evaluates proposed AI actions against your policy and returns ALLOW, REVIEW, or BLOCK — before anything hits your backend.

Try the structured demo Read API docs

No signup required to try the demo.

AI triggered refund ❌
Zorelan BLOCKED it ✅

Where Zorelan sits

User request→AI output→Proposed action→Zorelan policy check→ALLOW / REVIEW / BLOCK→Execute · review · block

Zorelan runs after model generation and before execution. It does not replace your models — it decides whether their proposed actions are safe to run.

Example — refund before delivery is confirmed

Correct-sounding reply. Unsafe action.

AI model output

“I have issued your full refund of $180.”

proposed_action

{
  "type": "refund_customer",
  "parameters": { "amount": 180, "currency": "AUD" },
  "reversible": false,
  "context": { "order_status": "delivery_unconfirmed" }
}

policy.controls.refund

auto_allow_limit: $100 AUD
absolute_review_limit: $1,000 AUD
require_delivery_confirmation_above_auto_allow_limit: true

BLOCK

Do not execute

Reason

Refund of $180 is above the policy's auto-allow limit ($100) and delivery is not confirmed.

Missing context

delivery_confirmed — required for refunds above the auto-allow limit.

Next step

Block execution. Request delivery confirmation, then re-evaluate.

See the full interactive flow in the structured demo →

For developers

One call, before the action runs

Send the proposed action and the policy it must satisfy. Branch on the verdict.

const decision = await zorelan.evaluateAction({
  user_request,
  model_output,
  proposed_action,
  policy,
});

if (decision.verdict === "ALLOW")  executeAction();
if (decision.verdict === "REVIEW") routeToHumanReview();
if (decision.verdict === "BLOCK")  blockExecution();

Gate execution on decision.verdict. Full reference in the API docs →

How decisions are made

Typed refund enforcement

For refunds, Zorelan enforces your typed policy.controls.refund (limits and confirmation) and returns ALLOW, REVIEW, or BLOCK — a repeatable check, not a model guess.

Honest policy matches

A refund decision cites the typed controls it actually applied. Free-text rules are explanatory context only and are never presented as enforced.

Missing-context detection

When a decision depends on information you did not send (such as delivery confirmation), Zorelan surfaces it as missing context instead of guessing.

Fail-safe by default

Account deletion returns BLOCK or REVIEW; subscription and CRM changes, and unknown actions, route to human REVIEW. No action is auto-approved on a caller-supplied boolean alone.

Prompt verification (multi-model comparison and trust score) remains available separately as the legacy /v1/decision path. The execution gate above does not rely on multi-model arbitration today.