The loops lab drafted two governance proposals and put them in a drawer: an amendment to the factory’s canonical gremlin model adding a new team pattern and a carefully scoped exception to its design principles, plus a placement note for the operating model. Drafted, never sent. Three days later the owner sat down to review them properly, and the review’s first move was the one this post exists for: it did not start by reading the proposals. It started by checking whether the world they described still existed.
Three days. Not three months. That is the uncomfortable punchline, so it goes first: the gap between “written” and “reviewed” that was enough to make a verbatim reading dangerous was seventy-two hours in an agentic factory. In that window, the specs the proposals cited had been restructured, once on the very day of the review, and the systems the exception was scoped against had kept moving.
What the reality-check found
The pre-decision verification re-checked every technical claim against the live repos rather than trusting the draft: the sandbox contract’s zero-guest-egress topology still matched, the adverse-condition test plan still matched, the neighbouring broker lab was confirmed still schema-only with no runtime, and the canonical target file was confirmed unchanged, with nothing pre-applied. The claims held, which is not the anticlimax it sounds like: “verified, still true” and “assumed, probably true” produce identical approvals and completely different audits.
The review’s real findings were structural. The exception’s cross-references were coupled to section numbers, spec §5, plan §2, in documents that had already been restructured once that same day, so the exception’s own pointers were a staleness time bomb. And the amendment was a factory-wide change to a canonical cross-project document, sourced from one lab’s design, which deserved naming as such rather than sliding through as lab housekeeping.
Accepted, with the fix built into the exception itself
The disposition: the amendment and exception accepted with changes, the placement note accepted as-is for the owner to apply within his own write authority. The changes are the reusable part. Section-number references were replaced with document-title plus section-name references, which survive renumbering. And the exception’s own review-trigger clause was extended: any future review must also re-verify the references if the cited documents have been renamed, restructured, or retitled. The exception now carries the instruction for checking its own assumptions, so the next reviewer does not need to rediscover this post’s lesson to apply it.
The part worth stealing
Open your oldest active exception, allowlist entry, or waiver and try to answer one question from its text alone: what did this assume about the world? If the answer is written down, check it. If it is not, that is the finding, and the fix is the template addition above, applied the next time the exception is touched. The velocity of agentic development does not change this discipline; it just shortens the interval at which it matters from quarters to days.
What Next
The Routing Policy Finally Learned Something
Also in Governance Without Killing the Joy: the same discipline pointed at rules instead of exceptions, and the counter that forces a policy to face its own evidence.