data-rule

What problem does it solve? Experiments often fail because of inconsistent or leaky data practices: unclear dataset provenance, train/test leakage, proxy labels that miss the target behavior, or sample sizes too small to trust. This Skill centralizes every data constraint an experiment must satisfy so results are valid and comparable across methods and milestones. ## Core Features & Use Cases - Provenance Rules: Enforces a preference order of existing, then adapted, then constructed datasets, and requires recording which was used. - Split and Label Integrity: Mandates explicit train/validation/test splits with deduplication and group-based separation, plus labels grounded in the dataset rather than another model's output. - Sample Size Floors: Sets minimum effective sample sizes — at least 50 examples for inference-time mechanism exploration or intervention, and roughly 100 for tuning or editing — unless the user specifies an amount. - Use Case: When designing a mechanistic interpretability experiment (e.g., steering coefficient sweeps), apply this Skill to guarantee each coefficient setting is evaluated on at least 50 held-out examples with no leakage. ## Quick Start Apply the data-rule skill to define the dataset, splits, labels, and minimum sample size for my mechanism exploration experiment.