What problem does it solve? Once main experiment results support a claim, researchers still need rigorous ablation studies that anticipate reviewer questions before paper submission, and designing them manually is slow and easy to bias. ## Core Features & Use Cases - Reviewer-Driven Ablation Design: An external LLM reviewer (via the llm-chat MCP) proposes component ablations, hyperparameter sensitivity tests, and design-choice comparisons from a reviewer's perspective. - Structured Ablation Plan: Normalizes reviewer output into prioritized tables with what each ablation tests, expected outcomes, run order, and GPU-hour estimates. - Feasibility Review and Execution: CC checks compute budget and code-change requirements, smoke-tests each ablation, runs them in order, and logs results in EXPERIMENT_LOG.md. - Use Case: After your mechanistic interpretability method passes result-to-claim validation, invoke this skill to generate a complete ablation plan, run config-only ablations first, and record negative results as findings. ## Quick Start Ask the agent to plan ablation studies for your validated method by describing the method and its confirmed claims.
