The governance kit

Most agents have no governing document.

The control plane assumes one exists. It versions it, scores it, gates its promotion and stamps it onto every run. That is the second half of the problem. The first half is that the document has to be written, and a compliance officer will not author a tool-use policy from scratch — neither will most engineers.

These four files are the first half. They are the same files a licensee receives, read from the repository rather than copied here, so the method described and the method shipped cannot drift.

The order it runs in

01

Extract

Read the agent's code. Its current prompt, every tool it can reach, every external write, what data it touches.

02

Interview

Derived facts first, as corrections rather than questions. Then the ones that cannot be derived, because they are policy and live in a person's head.

03

Draft

Fill the template. Anything unanswered is UNRESOLVED, never a plausible value.

04

Derive fixtures

Every clause produces at least one case. Every case names the clause it tests.

05

Review and score

A human edits. The suite runs. The gap between the document and the agent's actual behaviour is the work list.

06

Promote

A human promotes. The eval gate decides whether that is permitted.

The two rules

Never invent a policy judgement

Not what counts as irreversible, not who may approve, not the acceptable escalation rate, not the cost asymmetry. These are organisational decisions. An invented threshold reads exactly like a decided one on the page, and someone will follow it. A blank stops them. Unfilled means UNRESOLVED.

Every constraint gets a case

And every case names its clause. Traceability in both directions or it is not traceability. A constraint with no case is decoration that reads like governance in a review and constrains nothing at runtime. A case with no clause is an opinion someone encoded on a Tuesday, and when it goes red nobody can say whether the agent broke or the test was always wrong.