What problem does it solve? Turning a written experiment plan into running GPU experiments requires writing code, reviewing it for bugs, routing to the right mechanism family, and deploying runs — a slow, error-prone manual process. This Skill automates that bridge between idea discovery and the auto-review loop, producing initial results ready for iteration. ## Core Features & Use Cases - Plan-driven implementation: Parses EXPERIMENT_PLAN.md, EXPERIMENT_TRACKER.md, and FINAL_PROPOSAL.md to generate experiment code for sanity, baseline, main, and ablation milestones. - Mechanism-family routing: Routes the plan through mechanism-skills and experiment-tips catalogs to commit a mechanism family and adopt reproducibility conventions before any code is written. - Phenomenon-validation gate: Runs a phenomenon-validation milestone first with integrity auditing and a bounded hyperparameter-tuning retry budget before spending mechanism compute. - Cross-model code review and GPU deployment: Sends experiment code to an external LLM reviewer, then deploys runs in parallel with GPU partitioning, OOM auto-scale-up, and resume support. - Use Case: A mechanistic interpretability researcher finishes a refined proposal and says "implement experiments" — the Skill routes the mechanism family, writes and reviews the code, deploys runs across available GPUs, and collects initial results for the auto-iteration loop. ## Quick Start Tell the agent to implement and deploy the experiments from refine-logs/EXPERIMENT_PLAN.md for your research direction.
