Also called: false done, false success report
In plain words
A false completion claim is an AI agent saying it finished something — pushed to main, posted the update, raised the budget — when the destination doesn't show it. The agent usually isn't lying. It's sincerely wrong: a push failed on authentication after the command returned, an API accepted a post and the platform rejected it later, or the agent edited the wrong object.
What makes these claims dangerous is that everything downstream trusts them. Dashboards count them as completed work, alerts stay quiet, and increasingly people pay agents on the strength of them. The agent's own logs and traces look like success, because they record what the agent did and believed, not what the destination holds.
How QED Proof uses it
Catching false completion claims is the product. QED Proof never uses the agent's report as evidence: it waits until the work should be visible, reads the destination itself, and compares what's there with what was claimed. A false completion claim comes out as Failed (the destination was read and the outcome wasn't there by the deadline), Mismatch (something landed, but a required field differs, such as the wrong branch) or Late (it landed after the tolerance window).
A verifier that can't read the destination never calls the claim false. That's Couldn't check, so an agent is never blamed for QED Proof's blind spot.
Example
An agent reports: "Pushed a1b2c3 to main." QED Proof reads the branch's commit list on GitHub and finds no such commit. The receipt's verdict is failed with reason code not_found, and it carries the facts that let anyone re-check it.