What problem does it solve? When an experiment has many variations, manually reading every run's outputs and W&B metrics to write a cross-variation summary is slow and error-prone. This Skill automates that synthesis while keeping a human in the loop: the report is registered as pending review rather than finalized, so the experiment owner must approve it before it counts. ## Core Features & Use Cases - Cross-variation synthesis: Pulls every non-retracted variation (or a single variation, or a filtered corpus), reads outputs and real W&B metrics, and drafts a structured report with required sections including a mandatory "What didn't work" section. - Review-gated registration: Registers the report as a takeaways_report, variation_report, or research_report with pending review status; owner approval of an experiment-scope takeaways report is what unblocks experiment conclusion. - Findings and citations: Records one finding per variation for the experiment's running summary, checks claims against active research lessons, and attaches paper and prior-experiment citations. - Use Case: After running five ablation variations of a training experiment, ask the agent to distill the experiment; it fetches each run's W&B metrics, writes the synthesis with honest negative results, and registers it pending your approval. ## Quick Start Ask the agent to distill experiment <experiment-id> across all variations into a takeaways report pending review.
