Triage FAQ

Scope: the follow-up questions that survive Read a failing eval. If you have not read that page yet, start there. It decodes a result in one pass.

About the result

The eval says SKIP. Did anything actually run? Not that one. The reason is always printed underneath. There are four unrelated causes: a --deterministic-only run, no judge configured, an artifact that opted out, or an exhausted cost budget. A grader can also skip itself when the trace lacks the data it needs. None of them fail the build.

What is adheres-to-artifact (implicit)? The eval an artifact gets when it declares no evals of its own: did the session follow this file’s instructions? It is judged against the artifact’s full text. Nobody wrote it; you get it for free.

Why are there more evals than evals I recognise? Every artifact the session touched contributes, including project rules and any skill or subagent it used. Artifacts with no declared evals each add one implicit eval.

An artifact I expected is missing from the results. Check the coverage table below the evals. It lists every artifact the session referenced and whether the file was found. If it is not there at all, the session never referenced it.

What does a ○ in the coverage table mean? The reference could not be resolved to a file. It does not fail the run, and it does mean that artifact went unchecked. built-in agent (no definition file) is the expected case and needs no action.

About judged verdicts

The judge is wrong. What now? Check its observed line first. It cites what the judge claims it saw. If that citation does not match the session, the verdict is unsound and worth saying so to whoever owns the artifact. If the citation is accurate but the conclusion is not, the assertion is probably ambiguous, which is a eval problem rather than a session problem.

Votes were split. Which one counts? None of them auto-resolve. A verdict only lands as pass or fail when every run agreed and confidence cleared the threshold; anything else routes to REVIEW for a human. Split votes usually mean the assertion is ambiguous or the session genuinely is borderline.

error: is above zero in the votes. Is the result wrong? It is unreliable rather than wrong. One or more runs failed to produce a verdict. Errored runs count against consensus and can only push a result toward review, never toward a silent pass.

Why did it pass when I would have failed it? An auto-pass needs unanimity and mean confidence above the threshold, so a pass is not a weak signal. If you disagree, the assertion likely does not say what you meant. That is worth fixing at the eval, because it will keep passing otherwise.

About re-running

Can I reproduce this locally? Yes, and it is read-only: run never modifies a trace or the artifacts it evaluates.

npx moose-tracevals run <trace-file> --deterministic-only

I re-ran it and got the identical judged verdict. Is it cached? Almost certainly. Verdicts are cached on provider, model, prompt version, run count, temperature, the trace, and the eval plan. Add --no-cache to force fresh judgements.

Can I re-run without spending anything? --deterministic-only makes no model call at all. --provider mock exercises the whole pipeline offline but returns a canned verdict, so use it to check wiring rather than adherence.

The trace path in CI does not exist on my machine. Point artifact lookup somewhere real:

npx moose-tracevals run <trace-file> --project ./my-repo --deterministic-only

About ownership

Who fixes this? The coverage table printed the artifact’s file path, and that file is in version control. git log on it names the owner. Bring them the artifact name, the eval name, and for a judged verdict the vote split plus the cited observation.

I think the eval is wrong, not the session. That is a legitimate and common answer. An eval no session could satisfy is a defect in the artifact, and you may be the first person to notice. Raise it as an eval bug rather than changing your work to satisfy it.

The build is red and I cannot fix it right now. needs-review counting as a failure is configurable (failOnNeedsReview: false), and severity is per eval. Both are decisions for whoever owns the gate, though, not a workaround to apply under time pressure.