Hallucinations survive because the model defends its first draft. This flips the incentive: after drafting, the only goal is to destroy the draft. Every claim gets attacked. Whatever is still standing is the answer.
When to use
- Factual questions, research, technical claims, debugging conclusions.
- Anywhere a confident wrong answer costs more than a slow right one.
Behavior
- Draft the answer privately. Do not show it yet.
- Extract every load-bearing claim from the draft into a list: facts, numbers, names, API behaviors, causal statements. If the answer depends on it, it goes on the list.
- Attack each claim in turn. Attacking means evidence, not re-reading your own reasoning:
- Run the code, command, or query when one exists.
- Read the actual source, docs, or file.
- Search for disconfirming evidence, not confirming.
- Attack load-bearing claims from at least two angles.
- Verdict per claim:
- SURVIVED: attacked and held, evidence cited.
- SHAKY: could not verify either way. Must be labeled as unverified in the final answer, never asserted as fact.
- BUSTED: disproven. Correct it or delete it.
- Rewrite the answer from survivors only. Shaky items appear with an explicit flag and what would confirm them.
- If a busted claim was central, say the original answer was wrong and what changed. Never smooth it over.
Output
The verified answer with unverified statements flagged inline, followed by a short kill list: claims that got busted during the crash-out and what replaced them.