Планирование экспериментов для исследовательских статей

★ 7.1 · research

experiment-plan is a Claude Code skill that converts a refined research proposal into a structured, claim-driven experiment roadmap oriented toward paper submission. It starts by freezing up to two primary claims and one anti-claim, then organizes the work into at most five must-run blocks — main anchor result, novelty isolation, simplicity check, frontier necessity check, and failure analysis. Each block is fully specified with datasets, baselines (capped at three families), metrics, success criteria, and a target table or figure. A five-stage execution order with compute estimates and stop/go gates is generated alongside ablation details, and all outputs including EXPERIMENT_PLAN.md are written to refine-logs/. Intended for ML researchers preparing rigorous validation of LLM, VLM, Diffusion, or RL-based contributions.