santa-method

Use dual independent review: spawn two reviewer agents with no shared context, give both the same rubric and output, and require both to return PASS verdicts. Fix flagged issues and re-run fresh reviewers until convergence, capping at three iterations before escalating to a human.

Multi-agent adversarial verification uses two or more independent review agents to check generated output against an objective rubric. Because reviewers share no context, they avoid the correlated blind spots of self-review, and both must pass before output ships.

Define criteria with objective pass/fail conditions: factual accuracy, hallucination-free claims, completeness against the spec, compliance constraints, internal consistency, and technical correctness. Add domain extensions for content, code, or regulated industries. Vague rubrics produce vague reviews.

Skip it for internal drafts, exploratory research, and tasks with deterministic verification like build, lint, or test pipelines. It costs roughly 2-3x generation tokens per cycle, so reserve it for published, deployed, or compliance-sensitive output.

Use stratified sampling: run full dual review on 10-15% of the batch (minimum 5 items), classify failures by type, apply targeted fixes across the whole batch, then re-sample. This catches over 90% of systematic issues at 15-20% of full verification cost.

This is the infinite loop failure mode, where fixes introduce regressions or reviewers drift to new concerns. Cap iterations at three, instruct fix agents to change only flagged issues, and escalate to a human when the cap is exhausted.