exp-design

Experiment design with non-linear ablation generates a full test suite including ablation, sensitivity sweeps, and multi-seed main validation, then supports up to two re-planning iterations if certain factors prove marginal.

Non-linear ablation in experiment design evaluates method components iteratively, allowing the suite to automatically simplify and re-plan the method when specific ablated factors prove marginal or harmful to overall performance.

Benchmark and metric selection uses your idea page alongside relevant papers, concepts, and existing experiments within your research wiki to choose appropriate baselines, datasets, and evaluation criteria for validation.

Plan instrumentation for intermediate quantity analysis by logging gradient norms, loss decomposition, and attention or feature distributions during experiment design to support deep post-run analysis of why a method works.

No, this experiment design workflow must read your idea page and contextual wiki pages to generate grounded experiment designs, write design pages, update graph edges, and rebuild derived context artifacts.