A review pass has independent agents adversarially check each other's findings before any is reported, and a fresh model tries to refute each one. For one finding, that attempt neither confirms it nor refutes it. How does the report treat it?

The principle — an adversarial verification pass sorts findings into three states, not two, and the third one is where its honesty lives.

Why the key is correct — the behaviour is documented directly: the report lists that claim as unverified instead of counting it as refuted. The pass around it is the adversarial arrangement — independent agents adversarially review each other's findings before they are reported — and its grading rule is the same one that governs verification generally: a fresh model tries to refute the result, so the agent doing the work is not the one grading it. Read those together and the third state is inevitable. An attempt to refute can succeed, it can fail, and it can be inconclusive; only a pass that guaranteed it could reach every fact would be entitled to two outcomes.

Why the others are wrong — resolving every finding assumes exhaustiveness the pass never claims. Marking it refuted is the specific error the documentation rules out in the very sentence that describes the case, and it is the costly direction: a real bug that happened to resist reproduction disappears from the report as though it had been disproved. Treating agreement between reviewers as confirmation replaces the mechanism with a vote, and a vote among agents that each looked for a different class of issue is a measure of how visible a pattern is, not of whether the behaviour is real.

Remember this — unverified is a real answer: it means nobody could settle the claim, which is neither a bug confirmed nor a bug dismissed.

Sources — Anthropic's Claude Code dynamic workflows and best practices documentation.