ablation-planner

Ablation design starts by listing each removable or replaceable method component, then defining what each removal tests and the expected outcome if the component matters. This skill uses Codex to generate a prioritized plan covering component ablations, hyperparameter sensitivity, and design choice comparisons.

Reviewers typically expect ablations isolating each novel component's contribution, sensitivity tests for key hyperparameters, and comparisons against natural alternative design choices. The plan also identifies unnecessary ablations that would not add insight.

Run ablation planning after main results pass result-to-claim verification with claim_supported equal to yes or partial, or when a review loop identifies missing ablations. Planning before claim verification risks designing experiments for unsupported claims.

Component ablations take priority over hyperparameter sweeps, and config-only changes run before code modifications. If total estimated GPU-hours exceed budget, lower-priority ablations are proposed for cuts and re-prioritized rather than silently dropped.

The plan quality depends on accurate method descriptions and current experiment results being available in project files. It does not generate ablations for components identical to the baseline, and feasibility still requires manual compute budget confirmation.