experiment-design

Structure multi-stage experiments by applying a 4-stage progressive framework that guides initial implementation, baseline tuning, creative research, and ablations. This produces a complete design artifact with datasets, metrics, and hyperparameter grids for execution.

Plan baselines and ablations by generating an ablation matrix and hyperparameter grid across multiple datasets and seeds. The 4-stage progressive framework ensures systematic baseline tuning and creative research before final ablation studies.

Generate a hyperparameter grid and evaluation metrics by providing a research plan JSON or method name. The design process translates the input into an executable Markdown or JSON design artifact containing the required metrics and grid.

Yes, providing a research plan JSON or a method name produces a complete experiment design. The output satisfies the required design format, yielding a Markdown or JSON file ready for execution with baselines, datasets, and seeds.

The 4-stage progressive experiment framework includes initial implementation, baseline tuning, creative research, and ablation studies. It systematically structures the planning process to ensure comprehensive evaluation across multiple seeds and datasets.

You must define datasets, metrics, and seeds to translate the plan into executable design formats. Providing this information upfront allows the framework to accurately generate the ablation matrix and hyperparameter grid.