Meeting minutes classifier rework summary
Run directory: /root/meeting-minutes-evidence-model/runs/multitask-rework-20260811T150804Z/model_bundle

Headline test results:
- Evidence type: macro-F1 34.9%, micro-F1 44.6%
- Action state: macro-F1 49.7%, micro-F1 53.3%
- Signals: macro-F1 66.1%, micro-F1 78.0%
- Quality gate: macro-F1 66.6%, micro-F1 66.7%
- Quality defects: macro-F1 32.9%, micro-F1 34.5%

Quality-gate confusion matrix, rows are actual and columns are predicted:
['intervention', 'ready']
[[13, 2], [10, 11]]

Quality gate class detail:
- intervention: precision 56.5%, recall 86.7%, F1 68.4%, support 15
- ready: precision 84.6%, recall 52.4%, F1 64.7%, support 21
