EvalLayerTry the demo
The independent assurance layer for AI agents

Stop AI agents from making promises your business never approved.

EvalLayer checks quotes, refunds, discounts, delivery commitments, and customer replies against your policy before they ship. It returns allow, rewrite, require approval, or block, and keeps a signed record of the decision and what happened afterward.

Two surfaces, one engine

Judge the work. Gate the action.

One engine, standing in front of the send button: the same pipeline that scores completed deliverables also gates the next consequential action.

POST /authorizeLive

Authorize actions before they happen

A pre-action gate for outbound quotes, replies, refunds, and posts. Policy checks return allow, rewrite, require approval, or block, with a compliant rewrite when one exists.

Try the demo →
POST /evaluateLive

Evaluate completed work

Post-hoc verdicts on agent deliverables. Claim extraction, evidence matching, quality scores, and payout recommendations, each stored with a full audit trail.

Try it free →
Where it sits

Three layers, three different questions

LayerThe question it answers
Identity and accessCan this agent call send_email?
Security guardrailsIs the output malicious or unsafe?
EvalLayerIs this discount, refund, promise, or claim commercially authorized?

The agent may have permission to send the email. EvalLayer decides whether it may send this email.

Outcome architecture

The decision is only the beginning

Decision

What EvalLayer recommended: allow, rewrite, require approval, or block, signed and hash-chained.

Disposition

What the human or system actually did: sent it, used the rewrite, overrode the block.

Consequence

What happened afterward: the dispute, the complaint, the quiet success, with its own timestamp.

Adjudication

Whether a named reviewer ultimately judged the decision correct, recorded separately and revisable only by supersession.

EvalLayer keeps those layers separate so you can measure human disagreement, rewrite adoption, incidents, and adjudicated decision quality without treating every override as ground truth.

How it works

Every decision earns its verdict

Submit

Send a completed deliverable, or an intended action with the policy it must satisfy.

Extract

The pipeline pulls out every claim, promise, number, and commitment in the text.

Check

Each one is tested against supplied evidence or against your policy rules.

Verdict

A structured decision with reasons, returned in seconds and stored as an audit record.

Integrate in minutes

An API, not a dashboard ritual

quick startcurl
POST /authorize Authorization: Bearer sk_... { "action_type": "send_quote", "content": "...we can guarantee go-live by August 1...", "policy": { "max_discount_pct": 15 } } => { "authorization_id": "authz_9f2c...", "decision": "block", "risk_score": 0.95, "violations": [{ "rule": "forbidden_commitment", "severity": "hard" }, { "rule": "over_discount", "severity": "approval" }], "approved_rewrite": null }
Hard violations block with no rewrite. Softer ones come back as rewrite with a compliant alternative already re-validated. Call it at your send boundary, before the message ships: try it live. The quick start runs on a free key.
Production provenance

EvalLayer's engine first made financially consequential decisions in an autonomous agent marketplace.

276 completed-work evaluations 247 settled on-chain 88% pass rate The full history →

Also built on the same engine. Every verdict feeds a claims intelligence layer and a portable agent reputation system.

Ways to start

Three ways in, depending on where you are

Try Authorize
Free, no signup
  • Run your own message and policy through the live gate
  • See the decision, violations, and rewrite
  • About two seconds
Open the demo →
Founding design partner
No platform fee
  • Four weeks, real or shadow-mode traffic
  • Policy modeling built with you
  • Two slots
Apply →
Agent risk exposure audit
$5,000
  • Your historical agent conversations through the policy engine
  • The violations it finds, categorized and quantified
  • Tuned policy file and executive report
Ask about the audit →

Developer pricing for the Evaluate API →

Design partners

Shipping agents that quote, reply, or refund?

Two companies get the authorization layer early and shape the policy model with us.

Talk to us