What problem does it solve?
It prevents wasting GPU-months by forcing a plan-first, approval-gated experimental framework that can falsify your research claim with correct baselines, metrics, datasets/workloads, and statistics.
Core Features & Use Cases
- Plan-first experiment & framework contract: Produces a complete experiment + harness plan (decision rule, variables, baselines, datasets/workloads, ablations, statistical plan, and result→artifact map) before any framework code is created.
- Reference-aligned harness architecture: Specifies the end-to-end code organization (declarative configs, module interfaces, reproducibility, experiment tracking, and orchestration) plus a minimal walking-skeleton milestone.
- Walking-skeleton validation workflow: Builds an end-to-end pipeline on toy scale, validates determinism and metric logging, and only then proceeds to real method or real-scale runs.
Quick Start
Ask the skill to produce an approval-ready experiment-plan for your hypothesis, including baselines, datasets/workloads, ablations, and the exact decision rule, and wait for sign-off before it scaffolds any project files.