ablation-planner

Ablation studies isolate component contributions by removing or modifying specific parts of a model to measure their impact. You design ablations for components, hyperparameters, and design choices with clear success criteria to strengthen research claims.

Designing ablations for reviewer questions involves creating a structured plan that covers component-level, hyperparameter, and design-choice variations. Use project context and experiment logs to generate prioritized run orders with coverage rationale to preempt reviewer inquiries.

Structure an ablation plan by evaluating project documents and compute constraints to generate a prioritized run order. The output must include defined ablations, priority levels, coverage rationale, and total compute estimates expressed in a standardized format for automated parsing.

Yes, automated ablation planning works with limited compute by using resource constraints to generate a practical, ordered plan. It prioritizes ablations based on available budgets, ensuring critical experiments run first while tracking total compute estimates for reproducibility.

Trigger an ablation plan when main results pass the result-to-claim check and ablations are requested or triggered by an auto-review loop. This ensures ablation studies are initiated only after core claims are validated and require further evidential support.