Note ·
The annotations are the most valuable records
A taxonomy agents can read makes a company legible, and legibility has a known cost: the map simplifies, and where it simplifies the wrong thing, acting on it breaks what the map could not see. Scott’s examples are forests planted in rows and cities drawn on paper. A company’s version is smaller and everywhere: the integration two teams keep alive by talking, the approval that officially takes a committee and actually takes one phone call, the spreadsheet that is the real system of record for a process the CMDB describes incorrectly.
The design consequence is that an ontology for agents needs a first-class record type for divergence: a way for a person to say, about any node or edge, that the map is wrong here and this is how it really works, with their name on it. These annotations will be few, written by the people with the longest tenure, and they will be the most valuable records in the warehouse, because they are the only ones that encode what no system of record holds. An agent that reads the ontology without them is confidently wrong in exactly the places that matter; an agent that reads them knows where to slow down and ask.
Treating the annotations as the real content, and the exported structure as the scaffolding around them, is the inversion that keeps the whole project from becoming the high-modernist mistake with better tooling.