critical-council-reviewlisted

# Critical Council Review Convene a council of independent, domain-relevant experts to critically analyze a result the user has directed you to examine, then have a final executive weigh their findings and make a clear call on what to do next. The purpose is to establish the truth about the result's quality — what is sound, what is flawed, and what should happen next. **Application strength.** One eval run exists — 18 artifacts, 89 planted defects, one model, one author; results and raw reviews are in the project's `eval/` directory. What it measured: a council with seats in isolation earns its cost on precision, not on coverage. It found no more planted defects than a single careful pass on most artifacts, and roughly a quarter as many false ones. A council run sequentially in one context earns its cost on neither: under 5 points of recall for more than twice the output, and four major findings on an artifact that had none. A second run tested whether that recall figure was soft. Six independent authors wrote fresh artifacts, each briefed to plant defects *designed to survive a careful expert review*, each recording in advance which reviewer blind spot it believed it was exploiting. A single careful pass found **18 of 18**. Every prediction of a miss was wrong. Read that as a constraint on this skill, not a compliment to it: **on recall there is close to no headroom for a council to occupy.** The case for convening one rests on precision and on catching the shared premis