Your agent's tests pass.
Prove they can fail.
Corund is a merge check that proves an AI agent's tests can fail: it reverts the change and requires the new tests to go red, restores it and requires green, so a green check means the tests caught something, not that the agent gamed them.
base 434550de4a1b -> head 8f21c7aa93e0 | posture: OBSERVE (every check-run conclusion is neutral; BLOCK is a per-rule opt-in) check verdict conclusion C2 skip-audit PROVEN neutral C1 red-on-revert PROVEN neutral C1 red-on-revert: PROVEN tests/test_invoice.py::test_annual_discount: pass with the change; fail in call AssertionError — assert 240 == 200 on the reverted tree; rerun reproduced it — WITNESS PROVEN: 1 witness(es) failed by assertion on the reverted tree and pass with the change: tests/test_invoice.py::test_annual_discount Corund never reports a check it did not run. corund-action 0.1.0
How it works
-
01
Install
The Action runs in your own CI on every pull request: C1 red-on-revert and C2 skip-audit. Observe mode by default; blocking is a per-check opt-in.
-
02
Replay
Before you opt into blocking, replay reads your last 50 merged pull requests and writes a report of what would have been flagged. It posts nothing to GitHub.
-
03
Read the receipt
Each check posts its own verdict — PROVEN, FAILED, UNPROVEN, GAMED-SUSPECT, CRASHED or NOT-RUN — with the evidence lines behind it. Corund never reports a check it did not run.