santa-method

Use dual independent review: spawn two context-isolated agents that evaluate the output against the same rubric covering factual accuracy and hallucination-free claims. Ship only when both reviewers return a PASS verdict, otherwise fix flagged issues and re-review.

Multi-agent adversarial verification is a pattern where independent review agents, sharing no context with each other or the generator, evaluate output against an objective rubric. Independence breaks the shared-bias failure mode where one agent misses its own systematic errors.

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

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

If only one reviewer catches an issue, that issue is real and the other reviewer's blind spot is exactly the failure mode the method eliminates. Requiring agreement from both independent reviewers prevents single-reviewer gaps from letting defects ship.

The loop is capped at a maximum of three iterations. If reviewers still flag issues after the cap, the output escalates to a human reviewer instead of looping indefinitely, which bounds cost and surfaces genuinely ambiguous cases.