experiment-knowledge-harness

A file-based research memory for ML/DL/Quant projects. Sibling of project-knowledge-harness (which manages software-work memory: TODO.md / backlog/ / pitfalls/). This skill manages experiments: hypotheses, runs, conclusions, and the beliefs built on top of them.

The atomic difference from software work: in research, a failed experiment is a first-class deliverable (it saves future compute), results are only meaningful when comparable (ablation discipline), and numbers without provenance rot within weeks.

Five surfaces, separated by the question they answer: