What is an AI agent evidence record?

An ordinary log is kept by the party it describes. The people who most need to read it are the ones with the least reason to trust it.

When an AI agent books a payment, answers a customer or changes a record, the only account of what happened is usually the vendor's own log. The vendor can edit it, lose it, or send a screenshot of the good parts. A buyer's security reviewer, an auditor or an insurer has no way to tell which.

An evidence record is built for that reader. It answers three questions with something they can check themselves: whose agent is this, what was it allowed to do, and what did it actually do.

Five parts, each one checkable on its own.

• Signed papers (identity). A credential for the agent, signed with Ed25519. APIR publishes its public keys at /.well-known/jwks.json and its DID document under did:web:apir.ai, so a signature checks against a key anyone can fetch.
• A mandate. What the agent may do and may never do, written by its owner. With a mandate on file, "did it stay inside its authority?" becomes a question with an answer, not an argument.
• A log book (the Agent Black Box). Insert-only. Every entry carries a SHA-256 hash of its content and of the entry before it. Change one entry and every hash after it stops matching.
• An outside timestamp. Every night the head of the chain is timestamped by an outside RFC 3161 time-stamping authority. Nobody, APIR included, can backdate an entry without the timestamps disagreeing.
• A Trust Passport and an Evidence Pack. The passport is the agent's public, signed summary, graded from its record; an agent with no evidence is shown as Unrated, never given a made-up score. The Evidence Pack seals one agent's record for a deal, a security review or a claim, behind a public link.

No account, no login, no call to the vendor.

Paste an agent's name or passport ID at apir.ai/check-agent. APIR re-checks the Ed25519 signature against the published key, reads the mandate back and recomputes the log book from what is stored. Every public passport is listed in the registry. The signature also checks offline, against the JWKS file, with no APIR server involved.

AI assistants can run the same check. APIR's read-only MCP server, listed in the official MCP registry as ai.apir/verify, answers "is this agent who it claims, inside its mandate, and what is its record?" at https://apir.ai/api/mcp.

It proves the record is intact and dated. It never proves the agent was right.

• It proves the papers were signed by the key published under did:web:apir.ai, and that the log book has not been changed since each entry was written.
• For an agent that runs outside APIR, the log book proves what the agent reported and that nobody changed it afterwards. It does not prove the report was complete.
• The mandate is written by the agent's owner. The record shows whether the agent stayed inside it, not whether the mandate was wise.
• A passport reflects the agent's state as of its last assessment, and says when that was.

We don't promise results. We prove whatever happens.

Anyone who has to rely on an agent they did not build.

• Buyers. Security reviews now ask how an AI agent is controlled and what it did. A record they can verify ends the back-and-forth faster than another questionnaire answer.
• Insurers. AI insurers price from the system in front of them. The record shows them its history instead of a description of it.
• Auditors. A log they can recompute themselves, with timestamps from an authority the vendor does not control.
• Australian organisations. From 10 December 2026 the Privacy Act asks privacy policies to describe automated decisions that significantly affect people. See what the OAIC says.

What is an AI agent evidence record?

A history of what an AI agent did that someone outside your company can check without trusting you: signed papers saying whose agent it is, a mandate saying what it may do, and a hash-chained log book whose entries cannot be changed without it showing. APIR keeps one for each agent, signed with Ed25519 under did:web:apir.ai and timestamped nightly by an outside RFC 3161 authority.

How is an evidence record different from observability logs?

Observability logs and traces are kept by the team that runs the agent, for that team, to debug and improve it. An evidence record is kept for a reader outside the team: it is signed, insert-only, hash-chained and timestamped by an outside authority, so a buyer, auditor or insurer can check it without access to your tools and without trusting you.

What is a tamper-evident log for AI agents?

A log where any change to a past entry is detectable. In APIR's Agent Black Box every entry carries a SHA-256 hash of its content and of the entry before it, and the head of the chain is timestamped every night by an outside RFC 3161 authority. Edit or delete an entry and the hashes after it stop matching. Tamper-evident means a change shows, not that a change is impossible.

What is an AI agent passport?

A signed credential for an AI agent. APIR's Trust Passport is an Ed25519-signed W3C Verifiable Credential that binds the agent to its grade and its record. Anyone checks it at apir.ai/check-agent or apir.ai/verify-passport, or offline against the keys at apir.ai/.well-known/jwks.json.

Can someone check an APIR record without an account?

Yes. The check at apir.ai/check-agent is free, needs no account, and is never paywalled. The signature can also be checked offline against the published key, with no call to APIR.

Does APIR see my agent's prompts?

Not for an agent that runs outside APIR. Prompts and outputs stay on your side and only their SHA-256 hashes are sent to the log book, through POST https://apir.ai/api/v1/events.