A general-purpose discipline for analyses where being confidently wrong is worse than being openly uncertain. Forces a three-step routine — enumerate sub-claims, cite evidence, falsify — to reduce the failure mode where deeper reasoning produces more confident wrong answers. Trigger for non-trivial analytical questions where the conclusion can be subtly wrong: strategic decisions, comparisons, equivalence judgments, root-cause analyses, design critiques, regulatory or interpretive assessments, code review, hypothesis evaluation. Trigger also when you notice yourself constructing a persuasive answer without having verified it. Do NOT trigger for simple factual questions, creative work, or casual conversation — the overhead is not worth it there.
Cheap retrieval over an LLM-wiki — TF-IDF index with automatic refresh, structured wiki-link graph (target/heading/alias), and broken-link reporting. Built on top of the LLM-wiki paradigm (curated markdown notes with frontmatter). Use when an agent needs to find specific information without paying the token cost of full-file reads, traverse backlinks, or detect graph drift. Returns top-k snippets with metadata in ~500 tokens instead of ~10 000 for a full-file read.