Agent Skill Vetter — vet any agent skill or artifact before you install it

A gate that has never failed proves nothing. This one fails on purpose — and you can watch it.

The Agent Skill Vetter is the deterministic check we run before anything ships. Point it at an agent skill, a handoff note, a report or a table export and it returns PASS or CHANGES with the failing lines.

What it does

• re-derives the arithmetic it can find and flags the numbers that do not reconcile;
• requires a URL and a date on every cited source;
• refuses unlabelled numbers ("about 20%", "roughly 3x");
• refuses unsupported absolutes ("guaranteed", "always", "never");
• prints the failing lines, so the fix is obvious.

In a CSV/TSV, JSON or JSONL intake the values are read as data: sources, labels and absolutes are checked, but an equation written inside a cell or a JSON value is not re-derived - the report says what it checked and prints checked: 0 arithmetic claims.

What it is not: not an AI reviewer, not a style checker, not a linter for your code. It checks the claims-and-numbers class of mistakes — the ones that survive a careful read.

Runs offline: python3 gate.py artifact.md → verdict + the reasons. Python 3.8+, standard library only. It ships with a self-test that must fail a deliberately corrupted artifact (the negative control) — if that control ever passes, the tool is broken.

Why free: it is the slice of our Agent Company Kit that people ask for first, and a working gate is the best argument for the rest. Take it, run it, no strings.