Governing an AI system is not a single verdict at launch. It means repeatedly asking whether the thing is still trustworthy enough, whether whatever has drifted since last time is tolerable, and being able to show your reasoning to somebody who wants to argue with it.

Existing work tends to land on one side or the other. High-level frameworks describe principles but offer nothing you can actually re-run next quarter to see whether the answer changed. Narrowly metric-driven approaches give you numbers that move, without connecting those numbers to the governance decision anybody has to sign.

This methodology aims at the space between, with a formal way to represent and learn trustworthiness levels that stays light enough to use repeatedly. The word to notice again is contestable: the goal is not only a documented judgment but one somebody can challenge on the record, which is closer to how a credit rating works than to a certificate on the wall.