What problem does it solve? After an experiment's variations and runs finish, someone must judge what actually worked against what was pre-registered — pulling real metrics from W&B, execution logs, and attached reports rather than guessing from run statuses. This Skill performs that judgment in-context and optionally persists findings, lessons, and reports to the Chronicle record. ## Core Features & Use Cases - Metrics-grounded judgment: Enumerates variations and runs, pulls real numbers via chronicle.wandb_* mediation tools, execution_log assets, and attached reports, then classifies each variation as working, partial, or not_working against its pre-registered hypothesis. - Read-only first-class mode: "Just tell me what the results say" ends at the presented read; persisting findings, lessons, or reports happens only on explicit request. - Record-checked evaluation: Optionally validates claims via chronicle.evaluation_support before persisting, surfacing verdicts, assumptions, and matched failure modes. - Use Case: After an lr-sweep experiment completes, ask the agent to evaluate the runs — it produces a per-variation table of what worked, what didn't, and what's unexplained, then records findings and a review-gated takeaways report if requested. ## Quick Start Ask the agent to evaluate the results of experiment X and tell you what worked and what didn't against each variation's pre-registered hypothesis.
