What problem does it solve? Turning a research idea or hypothesis into a rigorous, executable experiment plan is error-prone: baselines get missed, ablations are incomplete, and success criteria stay vague. This Skill structures the entire experiment design process around claims, ensuring every experiment validates a specific claim with quantified pass/fail criteria. ## Core Features & Use Cases - Claim Scoping: Scopes claims across three dimensions (Target, Decomposition, Threats), reusing existing wiki claims or creating new ones when gaps exist. - Four Experiment Block Types: Designs baseline reproduction, validation, ablation, and robustness experiments, each with setup, metrics, seeds, and GPU-hour estimates. - Staged Run Order with Decision Gates: Orders experiments into stages (sanity check → baseline → validation → ablation → robustness) with explicit stop conditions between stages. - Optional Review LLM Pass: With --review, sends the plan to a review model that checks for missing baselines, unfair comparisons, and statistical rigor. - Use Case: Given an idea like "sparse-lora-for-edge-devices" in your research wiki, generate a complete experiment plan with wiki pages, graph edges, and a compute budget report ready for execution. ## Quick Start Ask the assistant to design an experiment plan for the idea sparse-lora-for-edge-devices with a 40 GPU-hour budget and review enabled.
