Phase 0: Exploration
Run all 6 pool models on every training query to collect execution traces.
Model Pool
Exploration Traces (click a row to expand)
Phase 1: Skill Handbook Learning
Discover skills from contrastive traces, profile agents, and refine the taxonomy.
Phase 2: Pareto-Optimal Handbook Selection
Evaluate candidate handbooks at different granularities and select the Pareto-optimal one on the validation set.
Phase 3: Testing
Deploy the selected handbook and evaluate on 500 held-out NQ test queries.
Routing Trace Examples
Curated examples showing skill-aware routing in action.