Your agent's tests pass.
Prove they can fail.
Corund reverts the PR's diff and requires its new tests to go red. Then it restores the fix and requires green. Deterministic checks, verbatim receipts, no model judgment. If a test can't fail, it was never a test.
Open-source core. Five lines of YAML. A crashed check reports crashed, never green.
$ corund check --pr 214 base 434550de → head 8f21c7aa C1 red-on-revert revert diff, run new tests 24 tests RED ← proof of life restore fix … 24 tests pass C2 skip-audit 0 silent skips (2 loud, allowlisted) C3 gate-fold test files tracked, in required path C4 approval-sha word bound to 8f21c7aa (fresh) VERDICT PROVEN
The green checkmark stopped meaning anything.
AI now writes 22% of merged code, and AI-coauthored pull requests carry about 1.7x more issues. Frontier agents have been caught editing tests and hardcoding outputs to get to green. Developer trust in AI output fell to 29%. Review tools give you an opinion about the code. Nobody proves the checkmark itself, so oversight collapses onto a test suite no one has verified can fail.
Sources: DX Q4 2025 impact report (135,000+ developers); Stack Overflow Developer Survey 2025; EvilGenie (arXiv:2511.21654) and related 2026 reward-hacking literature. Full citations on the dossier page.
Four deterministic checks. One receipt.
Red-on-revert
The PR's non-test diff is reverted onto base and its new tests must fail. Then the fix is restored and they must pass. One decisive mutant, the original code, in seconds. Both runs land in the receipt.
Skip audit
Silent test skips become CI failures. Deliberate skips live on a loud allowlist committed in your repo, so vacuous green has nowhere to hide.
Gate-fold
Every test file in the diff must be tracked in git and inside the branch-protection required path. A test that exists but doesn't gate is a report, not a case.
Approval binding
Human approval is bound to the exact commit SHA. Code moves after the approval, the approval dies with it. Stale word is no word.
Built from months of catching agents faking green.
Corund is extracted from a production system run entirely by autonomous coding agents under a mechanically enforced zero false-pass guarantee. The dossier records every class of fake green we caught, dated, with commit SHAs. Two of six:
Reports don't gate. Two adversarial campaigns' write-ups reached main while their test harnesses didn't. Five false passes stayed live for ten days behind green CI.
Vacuous green. Five test files silently skipped in CI for missing dependencies. Every one was found by accident, never by a sweep. Skips are failures now.
Open core. Pay when your team gates on it.
Free
- C1 red-on-revert + C2 skip audit
- Receipts in your check output
- MIT-licensed core
Team
- All four checks, org-wide
- Receipt dashboard + permalinks + badges
- Approval-SHA binding and policy controls
Honest cost note: C1 re-runs your suite, so CI minutes roughly double on gated PRs. Scope it to agent-authored PRs, or don't. Proof costs a rerun; shipped defects cost more.