See the input → responsibility → output diagram.
What it does for you: It is the one place that says which parts of the Factory may use an AI model, which model, how hard it should think, and why. Most parts never use one: recording, checking, running, isolating and showing work must give the same answer every time, and a model's opinion is not evidence. Where a model can help (proposing edits, and later reviewing them or drafting summaries and plans), the task names an ordered list of up to four approved models and effort levels, the check that must accept its answer, and what happens when there is no answer. The Factory takes the first model on the list that nothing rules out, and records which ones it passed over and why. To save cost it lowers the effort level first; a cheaper model is used only once a measured comparison shows it does the same job for less.

Meet the matcher. Which tool fits this task? Records which modules may use models and why; selection is not qualification. Meet the whole Plumb family.
How it links: Covers every module in the Factory, including planned ones such as the nine in the proposed discovery and post-development design, and a check fails if a new part arrives without saying whether it uses a model. It agrees with the Producers module: the same size limits, the same model for proposals. Nothing uses it yet: the Product gate does not consult it, and the planned delivery flow will ask it which model a task may run before anything is sent.
Honest limits:
- Nothing is qualified, so nothing runs. By your decision of 1 October 2026 the Claude command line is unavailable for the Factory's model roles, so every Claude model is marked unavailable. OpenAI models are meant to be reached through the Codex command line on your ChatGPT sign-in, never a Platform key; that route is only named so far, not built or checked, and the older Platform-key connection is kept but no longer used. The Codex command line has no dedicated setting for how hard the model thinks (only a general configuration override, which the planned set-up does not use), so the design's three checks by an OpenAI model (auditing the sorting, challenging proposals and reviewing release drafts) are listed as blocked: named, with the intended model and effort, but always refused until that route has its own check.
- The planned discovery sorting and the weekly synthesis name an OpenAI model through the Codex command line first, and keep their earlier Claude model second, marked unavailable. Those two OpenAI choices are refused until that route is checked and says how hard the model thinks. The weekly web search keeps its separate Claude set-up with a search tool, which stays unavailable. Sorting is a guess until it is measured on the same work, and the Factory refuses to run it until its own measurement is recorded. Nothing runs, and neither task can be admitted by the schedule until the schedule records which model was chosen. Each choice follows a default that you have not yet signed.
- A list lets you add any provider by describing its route and models, but that only lets the Factory pick it. A new provider still needs its own check before it can be trusted, and its own account before the schedule can admit it. Checking by another model is only as independent as the two models are: both are now from OpenAI, so they may share mistakes.
- Only proposing edits has a finished question-and-answer format. Review, summaries, briefs, planning and every planned discovery and post-development task each need their own before they can run.
- The choices of model and effort are reasoned guesses, each with a measure that would retire it; none has been measured yet. Model prices and limits are a dated copy, not live.
- It decides nothing by itself and grants nothing: a model's answer is only ever a proposal that other checks accept or reject.
Status: Accepted: independent code review (Astra high) passed on 29 September 2026 in round 2; the 1 October 2026 priority-list change was merged the same day after an Opus 5.5 review and a Fable 5.1 review both passed. Acceptance and evidence.
The draft standalone adoption guide explains the private policy entry point and refusal proof. The distribution was accepted locally on 1 October 2026 (owner-designated acceptance reviewer, after independent Opus 5.5 and Fable 5.1 reviews).
Shared value implementation: common contracts. Owning errors, limits, stored identities and authority remain as specified above.
