Measurement Investigation

A system can be fully tested, ship daily, and still have measurement that means nothing. Tests check that the code does what its author intended. They never check that the metric measures what the dashboard claims, that the number had inputs, or that the difference between two runs is bigger than the noise. Those gaps are invisible to unit tests, type checkers, and ordinary code review — which is why they survive for years in codebases that look immaculate.

This skill asks one question, of every quality/accuracy/score number the system reports or a human trusts:

What does this number require to compute, on how many real examples was it computed, and what does it NOT tell you?

Let SKILL=/path/to/this/skill — the directory holding this SKILL.md. Commands run from the repository under audit, not from the skill directory. Needs Python 3.11+; the scripts are stdlib-only and make no network calls.