santa-method

Use dual independent review: spawn two reviewer agents with no shared context, give both the same rubric covering factual accuracy and hallucination checks, and ship only when both return PASS. Single self-review misses errors because the reviewer shares the generator's biases.

Multi-agent adversarial verification runs two or more independent review agents against generated output, each instructed to find problems rather than approve. Both must pass for the output to ship, and failures trigger a fix-and-re-review loop with fresh reviewers each round.

Skip it for internal drafts, exploratory research, and tasks with deterministic verification like build, lint, or test pipelines. It is designed for output that ships to users, carries compliance risk, or has elevated hallucination risk such as statistics and API references.

Cap iterations at a maximum of 3 rounds and escalate to a human when the cap is exhausted. Use rubrics with objective pass/fail criteria to prevent reviewers from flagging subjective style preferences that never converge.

Yes, use stratified sampling: run full dual review on 10-15% of the batch with a minimum of 5 items, classify failures by type, apply targeted fixes to the whole batch when patterns emerge, then re-sample until a clean sample passes.