Glossary: verifying AI agents' work

A

Agent claim: A structured statement from an AI agent that it performed an action on a named target — the untrusted input QED Proof verifies.
Agent reputation: A track record for an AI agent computed only from verified outcomes rather than reviews or self-report — the "credit score for agents". QED Proof's score isn't built yet.
AI agent verification: Independently confirming that an AI agent's claimed work actually happened, using evidence the agent didn't produce — and making the result checkable by others.
Append-only log: A log that can only grow — entries are added, never edited or removed — and that can prove it. Every QED Proof receipt is written to one.

Content fingerprint: A hash of normalised content that proves what was said without storing what was said — how a QED Proof receipt checks content while keeping none of it.
Couldn't check (unverifiable): The verdict given when a verifier can't determine what happened — never evidence for or against the agent, and never silently upgraded to Verified.

Destination verification: Checking an AI agent's work by reading the system the work was meant to land in, rather than the agent's own logs or report.

Ed25519: A fast, widely used digital-signature scheme (EdDSA over Curve25519, RFC 8032) with 32-byte public keys and 64-byte signatures. QED Proof signs every receipt with it.
ERC-8004: A draft Ethereum standard for trustless AI agents, with registries for agent identity, reputation and validation. QED Proof plans to post its verdicts to the validation registry; that isn't built yet.
Ethereum Attestation Service (EAS): An open protocol for recording signed statements, or attestations, on Ethereum and its layer-2 networks against registered schemas. QED Proof uses it to anchor its log on Base Sepolia (testnet).

False completion claim: When an AI agent reports that a task is done but the outcome never happened at the destination — the failure QED Proof exists to catch.

Inclusion proof: The short list of sibling hashes that proves one entry is in a Merkle tree with a given root — how a receipt proves it's really in the log.

JSON Canonicalization Scheme (JCS): RFC 8785 — a rule for serialising any JSON value to exactly one byte sequence, so that it can be hashed and signed reproducibly.

Merkle tree: A tree of hashes in which each parent hashes its children, so one root hash commits to every item beneath it — the structure behind QED Proof's receipt log.

On-chain anchoring: Publishing a log's root hash to a public blockchain so it's fixed in time and can't be backdated — QED Proof anchors its receipt log on Base Sepolia (testnet) through EAS.

Proof of Agent Work (PoAW): The technical name of QED Proof's open receipt spec — the format, verifiers, node and checker, published under Apache-2.0. Not a token, and unrelated to proof-of-work mining.

Reason code: A machine-readable code on a verdict that says exactly why it came out that way — for example not_found, target_mismatch or permission_denied.

Trust levels: The four levels of assurance a Proof of Agent Work receipt can reach — Signed, Anchored, Attested and Web-proven. QED Proof receipts reach level 2 today.

Unclaimed change: A change at a destination that no agent claim explains — an agent that acted and didn't say so. Detecting them is part of QED Proof's pipelines design, which isn't live yet.

Verdict: The outcome of verifying an agent claim — Verified, Late, Mismatch, Failed, or Couldn't check — decided from what the destination shows.
Verification receipt: A signed, tamper-evident record of one verification — the claim, what the destination showed, and the verdict — that anyone can check without trusting the issuer.
Verifier profile: The versioned, published rules for checking one kind of action — what to read at the destination, how to compare it, and how long to wait.