What broke

The station keeps its rules and habits in one manual, and agents read it before they work. One page explains how a review gets handed back: an author who has finished the changes a reviewer asked for tells the reviewer so.

The station's forge client had no verb for posting a review until 2026-10-28. So the handback ran through a substitute: a verdict filed as a shared artifact rather than as a review on the pull request. That substitute existed for a real reason, and the manual recorded it.

When the verb landed, the paragraph about the handback kept the substitute as its default and said a verdict is usually an artifact rather than a forge review. The same file, a few lines above, already described the review verb. The page contradicted itself, and the older half taught a habit whose reason had gone.

The belief was that a written workaround describes the system. It described the system on the day it was written.

How it showed

Nothing failed. An agent reading the lower paragraph would file a verdict as an artifact, and the reviewer would receive a link instead of a review on the pull request, so the verdict would not appear in the record that a merge guard reads.

The contradiction was visible to anyone reading the page top to bottom, which is how it was found: the section above and the paragraph below described two different protocols for the same act.

What was watching

Nothing compares the manual with the tooling. The manual is prose, its pages are checked for rendering and links, and no test asks whether a sentence is still true.

The station's own record made the drift legible: the review records on the forge show verdicts arriving as reviews after the 28th, and the artifact route leaving fewer traces each day. Reading two weeks of both is what showed the paragraph was describing a past system.

The detector was a person reading the manual for the current procedure and finding the old one described as the normal case.

Where it sits

This is a documentation-drift case.

The family is comprehension because a reader of the manual forms a model of the procedure from it, and the model the paragraph produced was of a system that no longer existed. It is evolution because the drift was produced by a capability appearing while a workaround stayed written down.

Documentation drift is exactly the missing sensor, and this station has no instance of it: the nearest thing is a link checker. Contract tests could hold the narrow part that is testable: a command the manual tells agents to run exists, and a command the tool has is described somewhere.

What we did

The paragraph now describes what the tool does, and the drop it records is kept, because the reason an author tag exists is still the reason the substitute existed. The two halves of the page agree.

What now fails loudly

Nothing, and the case says so. The manual has no test that a sentence is still true; a page can describe a retired procedure for as long as nobody reads it.

The control this class wants is a narrow one and would have caught this instance: every command the manual shows is executed by a check, and every verb the tool advertises appears in the manual. That check does not exist. What exists is the review of the manual change that the correction itself went through, which is the same sensor that found the problem.

This case is filed partial for that reason. The paragraph is fixed; the class is not.