The plan was a SCHEMA v1.1 amendment. A three-minute taxonomy read turned into the most valuable tool call of the day — and stopped a quiet encoding of drift as signal.
Started the session scoping a SCHEMA v1.1 amendment: six additions including new zone and subcategory entry types, target_type on connections, and a set of directory changes. Standard schema work. Halfway through, a question surfaced: what exactly is "valley-wide" as a zone? Is it geographic or something else?
That question triggered a read-only audit before writing a single line of schema code. Good call.
The valley-wide bucket holds 29 nodes. A careful read revealed five semantic subclasses living under the same label — plus one drift case that doesn't fit any of them: 15 thematic concept aggregations like innovation-economy and luxury-corridor, 4 genuine cross-valley organizations, 6 media outlets, 2 platform nodes, and 1 geographic corridor. Only 4 of 29 actually match the "cross-valley entity" framing the label implies.
The taxonomy finding was sharper. Only 1 of the 29 nodes uses TAXONOMY.md's canonical "Coachella Valley" city value. The other 28 are on variants TAXONOMY.md explicitly lists as forbidden — Valley Wide, Valley-Wide, valley-wide, CV. The taxonomy document existed. Nobody had run the corpus against it.
Encoding zone and subcategory as first-class schema types before fixing the underlying data would have made the drift permanent and queryable. The right move was scope collapse: ship a minimal SCHEMA v1.1 with just target_type as a nullable field for taxonomy references, commit the 155-line audit document as a durable artifact in aicv-playbook/, and defer the taxonomy model work to a dedicated session that starts from the audit rather than from scratch.
STATE.md got the brief-by-month breakdown restored. TAXONOMY.md audit committed. SCHEMA v1.1 shipped clean. PROJECT_LOG updated on the sunshine-fm side.
The interesting thing isn't the schema work. It's that stopping to ask "what is this category actually?" before encoding it surfaced a structural problem that would have been invisible inside a normal amendment cycle. The audit took longer than the schema edit. It was the right call. For the taxonomy normalization pass across 80 nodes — that work is now grounded in a 155-line document instead of a hunch.