Routing and prompts — choosing and recording analysis

Verification status — 6 October 2026

The effective-prompt audit is the current evidence for the selected analysis paths. Stored instructions, code-built instructions and output checks differ by path; a route name alone does not identify everything sent to a model. The August inventory and September notes below preserve earlier observations, including conflicting claims about explicit routes and default handling. Their counts and completeness claims are not current certification. The approved prompt standard remains the requirement.

Routing chooses how to analyze incoming material. The following inventory was recorded on 17 August 2026, with the September observations dated separately. It does not establish the current instructions or behavior of every analysis path.

The two layers recorded in August

Layer one — by source — is live. Sixteen active routes pick a prompt by where the piece came from: each social platform, each ad library, landing pages, the newsroom, and so on. This is the router the pipeline actually uses.

Layer two — by what the piece is about — exists but is switched off. Twelve rows, every one inactive, keyed on the old mixed content-category carve. The family reconciliation is what re-keys or retires this layer: the recommendation keeps routing on facts knowable at ingest and keeps the family out of prompts entirely.

Prompt inventory recorded in August

Forty-three prompt versions exist; nineteen are marked production. The system half of each prompt is hand-written per route; the user half is shared template text. Two honest gaps, found by probing:

• Version stamps are mostly missing — eleven of the nineteen production prompts carry no schema-version key, so "which vocabulary was this classified under" is often unanswerable. The ruled future — every classification stamped with its taxonomy version — needs this fixed.
• Prompts are not yet generated from the registry. The standing ruling says every choice list an analyzer sees is generated from the governed registry; today the lists are written by hand and kept in sync by discipline. This is the single highest-leverage build in the routing area.

Earlier drift observations — August and September

• The router never looks at what a piece is about. Every live source has a layer-one route, so the switched-off second layer is never reached — by design, since June. That means nothing in the pipeline decides whether a piece is creative work or supporting material: a model card, a research paper and a launch film from the same site all take the same road. About 290 pieces sit in the research and model-documentation categories today; no rule treats them differently. (Re-read on 2 September 2026: the earlier wording here said those categories "have no route", which described the dead layer, not the gap.)
• Three source types in live use — vendor portfolios, the images cut from landing pages, and embedded videos — have no routing row, so they ride default handling (re-probed 2 September 2026: exactly three).
• A dead twin of the routing table sits in the database, structurally identical, queried by nothing — leftover from an earlier design, worth retiring.

What the family ruling changes here

Under the taken-provisionally mechanism, nothing about routing stores a family: the router keys on source and format at ingest and on type after classification. The dormant second layer either re-keys to facts that exist at ingest or retires. Prompt generation from the registry is what makes the ratified vocabulary actually reach the analyzers.

The approved prompt standard

The same vocabulary definitions should serve people reviewing the taxonomy and the prompts that classify work. Choice lists come from that shared vocabulary. Prompts must explain what to choose and how to choose it, with detailed examples that distinguish easily confused meanings. Shorter instructions are not a goal when they make the classification ambiguous.

These are approved requirements. They do not establish that every current route, composed prompt or validation step follows them. The implementation audit must compare the actual instructions and allowed outputs with the shared definitions, then check the classifications and review experience on real work.

Where to look next

• The content pipeline — where the routed prompt runs.
• The schema view — the full pipeline walk with the reconciliation options compared.