Frozen review routes for evaluating generated companion notes. The existing 25-note review and seven-note checking set are retained in the linked reports.
Score each factor as supported, related-but-misleading or unsupported, and record missed broad facets separately. Review dates, polarity, procedural constraints and specific names explicitly.
Reports: selection experiment, evidence gate, and experiment ranking.
This is an exploratory set, not blind gold-standard data. Add every production false match as a new labelled case before changing the reranker.