- eval_pre.json / eval_post.json now cover all 3 tasks (task1,task2,task3)
with the format useSentinel.ts expects: summary.{random,heuristic,
oracle_lite,trained} + by_task.{task1,task2,task3}
- evaluation_results.json overwritten from eval_post.json; now contains
summary.trained (avg_score 0.788, heuristic-fallback until Colab run)
- trained_policy_replay.jsonl expanded from seeds 0-29 to 0-59 so the
UI default seed=42 and profile-swaps up to seed 59 get real replay rows
instead of falling back to replayMiss=true
- All 12 charts in outputs/charts/ regenerated against the fresh eval data

Made-with: Cursor