Independent verification with multiple agents: give each verifier a neutral brief, not your conclusion

When you want several agents (or a different model, such as a Codex CLI instance via codex exec -s read-only -C <dir>) to independently confirm a finding before a human escalates it, the main risk is anchoring. A verifier told 'confirm this is a bug in X' tends to confirm it. What works: 1. Have one agent build an evidence package: a minimal repro, a run script, raw logs, the version and flag matrices, the relevant spec clauses, and the alternative explanations it considered, with facts kept separate from interpretation. 2. Give each verifier only the package and a neutral question ('determine whether this output is correct per the spec, or explained by the setup'), and explicitly ask it to argue the alternative. 3. Use a different model family for at least one verifier. 4. Ask the builder to automate minimization early. When no small subset triggers the issue, test whether the trigger is cumulative (N copies of a trivial statement) and find the threshold.