Corund vs a model's judgment
Ask a language model whether a test is any good and the answer can move between runs; Corund never asks, because each of its four checks is a deterministic comparison, so the same evidence produces the same verdict — and when the evidence itself will not reproduce, Corund reports that instead of guessing.
| A model's judgment | Corund | |
|---|---|---|
| What decides | A model's answer about the code | Two runs, compared — no model judgment reaches a verdict |
| The same pull request, twice | The answer can move with the model version, the prompt, or the sampling | The same evidence produces the same verdict, on any runner and for any reviewer; a red that does not reproduce on the rerun never counts as a witness, and when it was the only red and nothing else was refused the verdict is UNPROVEN-flaky, not a guess |
| How it finds fake green | It reads the diff and forms an opinion about it | red-on-revert takes the fix out and requires the new test to go red; skip-audit refuses the vacuous green of a silent skip; gate-fold catches the gamed checkmark of a required check that stopped gating; approval-SHA binding fails an approval older than the code |
| When it cannot decide | It answers anyway | UNPROVEN with the reason, or CRASHED — never folded into a pass |
| Whether it can be required | An answer nobody can reproduce is hard to require | A verdict anyone can re-run from the two commit SHAs on the receipt |
A gate has to mean the same thing on Thursday that it meant on Monday, which is the whole reason no model judgment decides a Corund verdict.
For what an opinion catches that a comparison structurally cannot, read the LLM code reviewers comparison.