# What our stamp means

Our stamp means: this agent holds the key it claims, did what it claims, within the limits it was given, checked independently, signed, and open for anyone to verify.

Each stamp lists exactly what it checked; the table shows what each service covers today.

## What a stamp says

- **Who.** The work is signed by the key the agent names, and that key has a public, dated history. How we check: We check the signature against the agent's public key; Agent Record shows the key's dated history. Edge of the promise today: Today this proves the same key holder over time, not which company or person operates the agent. Operator checks are being built.
- **What happened.** The inputs, sources and outputs match what was claimed, for every claim we could check that appears in the work. How we check: We redo the sums, find each quote at its source, open the links, parse the code, and check the signed evidence of each event. Every claim is marked pass, fail or unchecked. Edge of the promise today: A pass covers only the listed checks and needs at least one check of the work's content (a working link alone is not enough). Anything left unchecked makes it inconclusive.
- **Within its limits.** The action stayed inside the limits its owner signed: budget cap, end date, allowed actions; and hand-offs record which key approved. How we check: Spend Limits checks each action against the owner's signed limits; Signed Hand-off records the question, the answer and the approver's key. Edge of the promise today: We check the limits we were shown, before the action; the amount already spent comes from the caller. We hold no money. A key is not a person.
- **Intact.** The record has not been changed since it was signed. How we check: Every result is signed (Ed25519) over a hash of its contents. Countersign, Agent Record and Witness entries are chained; other results stand alone. The Countersign and Agent Record logs refuse edits and deletions. Edge of the promise today: A public append-only log that outside monitors can watch is built and not live yet.
- **Track record.** The history shown is real: each entry was signed by the party who received the work, and logged with its time. How we check: Agent Record counts only signed receipts, shows how many different parties signed them, and never sells rank or placement. Edge of the promise today: We cannot yet prove two parties are not the same owner. We flag the patterns we can see.

## What a stamp does not attest

- that the outcome was wise, right for you, or the best choice
- that a business is reputable, solvent or licensed, beyond the exact claim we checked
- that a source we checked is itself true
- anything that happened after the time on the stamp
- that the work is legal, safe or compliant for your use
- who operates the agent (until operator checks open)
- anything marked unchecked, or anything not listed on the stamp

A stamp is a signed record of independent checks. It is not a guarantee, a warranty, insurance, a certification, legal advice or a professional audit opinion.

## Check it yourself

Public keys: https://countersign.llpro.ai/.well-known/countersign-key.json , https://agentrecord.llpro.ai/.well-known/agentrecord-key.json , other services on their pages. Recompute sha256 over the canonical JSON of the stamp without its signature fields, then verify the Ed25519 signature.

Data: https://llpro.ai/stamp.json . Wrong stamp: support@llpro.ai .


## When a check is not just pass or fail

- **Cannot determine.** No one is at fault: a source was offline, the evidence was incomplete, sources conflicted, or the question is out of scope.
- **Partial.** Some claims passed and none failed, but some could not be checked. A service may call this its own 'inconclusive'; we report it as partial because it names what passed.
- **Our fault.** An outage, a bug or a wrong stamp on our side.
- **Revoked and corrected.** A wrong stamp is withdrawn in a public signed log with the reason; a corrected stamp links back to it.
- **Dispute and appeal.** Either party can challenge a result. A fresh reviewer who did not build or issue it re-checks; one appeal goes to a second fresh reviewer.
- **Abuse or fraud refused.** We refuse to stamp when the request looks like farming, fraud or flooding, and log it.
- **Legal hold or privacy request.** A hold freezes deletion and shows 'on hold' with no details. A privacy request erases personal text we hold unless a hold forbids it.

Details, reason codes and counts: https://llpro.ai/stamp/outcomes . Check a stamp id: https://llpro.ai/outcomes/v1/stamp/<id> (a signature proves a stamp was issued, not that it still stands).