Also called: claim
In plain words
An agent claim is the agent's own account of what it did, in a form a machine can check. It names an action (what kind of thing happened), a target (where it happened) and the parameters that pin it down (which branch, which commit, which URL). It's the difference between "I think I'm done" in a chat window and a precise statement that can be proven true or false.
A claim is untrusted by definition. It says what to look for; it's never evidence that the thing happened.
How QED Proof uses it
Agents send claims to the claims API today, with one authenticated call (POST /v1/claims) using an API key from the console. A Python SDK, an MCP tool and a Claude Code hook are planned, not built. Each claim names an action string such as github.commit.push, and its parameters are checked against that action's verifier profile. Unknown parameters are rejected, because they're copied into public receipts.
The claim is then bound into the receipt by its digest: claim_digest is the base64url SHA-256 of the canonical (JCS) form of the claim. Anyone holding the receipt can confirm it was about exactly this claim, as submitted.
Example
POST /v1/claims
{ "agent_id": "release-bot",
"action": "github.commit.push",
"target": "Nuraveda/qed-proof",
"params": { "branch": "main", "sha": "a1b2c3…" } }