The rules are the output. This is the work that produces them and keeps them current: the calls, the extraction, the approvals, and the eval that scores every draft. Live sample workspace for a pursuit team; every name, deal, and number is fictitious.
Scored by an automated eval against the skill file: which rules applied, which flags raised, which numbers re-run, whether every claim carries an owner. The same eval runs on every draft, so the delta is tracked over time, not demonstrated once.
A file is a snapshot. The engine is the weekly calls that fill it, the approvals that keep it honest, and the calibration that checks it against what your team actually sent. The work is continuous, and it is what keeps the file from becoming a document nobody trusts.
They do not write things down. They react, and they talk. Iris takes the 30-minute call, asks the follow-up, drafts the rule, and nothing goes live until the person whose name is on it taps approve.
Every draft your AI produces is scored against the skill file, generic and Metis-powered side by side. The number at the top is that score, tracked over time. It is the one number, agreed up front, made visible.