Планирование ablation-исследований для ML-статей

★ 7.1 · research

ablation-planner is a Claude Code skill that systematically designs ablation studies to validate the contribution of individual components in an ML model before paper submission. It activates when main results pass the `/result-to-claim` check with `claim_supported = yes` or `partial`, or when a reviewer flags missing ablations. Codex (gpt-5.6-sol, reasoning effort xhigh) takes the reviewer's perspective to design the plan: isolating each novel component, testing hyperparameter sensitivity, comparing alternative design choices, and assessing coverage of likely reviewer questions. Claude Code reads `research_contract.md`, `EXPERIMENT_LOG.md`, and server configs to verify compute budgets, prioritize runs, and implement experiments — config-only changes first. Every ablation requires a defined `what_it_tests` and `expected_if_component_matters`; all results, including negative ones, are recorded in `EXPERIMENT_LOG.md` and `findings.md`.