The Expert Review Loop

Knowledge extraction is not a single pass. It is a loop: extract, synthesize, validate, extract again. The review loop is structured because unstructured review produces vague corrections — "this is not quite right" — that do not improve the artifact. Structured review produces specific corrections — "this step is missing, this condition is wrong, this scenario would fail here" — that close the extraction gap systematically.

1. Review Session Structure Present the knowledge artifact in sections, not as a whole. Ask the expert to review each section before seeing the next. Whole-artifact review produces impressionistic feedback. Section-by-section review forces the expert to engage with specific claims and produce specific corrections. Budget forty-five to sixty minutes per review session.
2. Correction Protocol For each correction, probe: "What should this say instead?" "Is this wrong in all cases, or only in some cases?" "Can you give me a specific example where this would fail?" The correction is not the information — it is the prompt for the information. A correction without elaboration is not actionable.
3. Residual Knowledge Probing At the end of each review session, ask: "What does this artifact not capture that someone following it would need to know?" "What is the most important thing I have missed?" "If someone used this artifact and produced a bad outcome, what would most likely have gone wrong?" These questions are designed to surface what the artifact leaves implicit.
4. Review Cycle Count Plan for two to three review cycles as the baseline. After each cycle, the number of corrections should decrease. If corrections increase after cycle two, the extraction methodology has a problem — the knowledge is being destabilized rather than refined. Three cycles with declining correction rates is the target. More than three cycles with flat correction rates indicates a deeper extraction problem.