Synthesize reusable benchmarks from interaction traces with feedback
scores. Mines high-quality traces, groups by query class, picks the
best reference per class, and exposes results through DatasetProvider
and LLM-judge Scorer so EvalRunner can evaluate against personal
workflow patterns.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>