Four modes + bite and test

Note · KBMill · control loop · corpus construction · frontier models

A frontier model can build a typed, provenance-aware corpus fast — if publication stays gated and work stays bite-scoped. The plant does not need to get larger first. What it needs is a corpus-construction control loop: four modes, one brick at a time, acceptance test before the next bite. That turns the model from a document summarizer into a supervised corpus builder.

Evidence lives on the public shelves: AI legislation bricks and executive-order bricks — separate classes, machine catalogs, hole cards, no silent melt of companions. The lesson is the loop, not the topic.

Unbounded “build me a KB”

Model invents scope and pipeline
Five mills before any check
Research auto-publishes
Cleanup deferred to the end

Fast, then a mess. Errors get large.

Bite → test → next

Plant sets one-brick grain
Same checklist every Manufacture bite
Research proposes; does not ship
Fail → fix this brick; do not proceed

Errors stay small. Cleanup stays small.

Four modes (same plant, different SOP)

Same stations. Different job, tool budget, and stop. Do not run Research tools while you think you are Manufacturing, and do not invent bricks to fill an Analyst matrix cell.

Mode	Job	Stop when
Manufacture	Build one known brick from supplied material (prep → mill → pack)	Acceptance test passes for this brick — only then start the next
Research	Discover candidates under a brief; propose membership; optional staging mills	Brief coverage met or explicit “enough for v0” — no auto-publish
Librarian	Families, versions, adjacency, rescinds links, hole cards	Graph and catalog consistent; versions not silently merged
Analyst	Compare accepted bricks; cite version + section	Findings cite bricks; synthesis labeled not statute

Mode locks

Research must not silently ship. Propose → accept membership → then Manufacture.
Analyst must not invent bricks to fill a cell — hole or unavailable only.
Librarian owns version families (companions, IH≠RH, EO rescinds) so Manufacture does not double-count.
Manufacture stays boring — one stock shape, closed tools, residual honesty.

Plant pacing: one brick → test → next

The number of Manufacture bites is not invented by the model. The plant already handles one brick at a time ((company corpus, same grain). The control loop makes the agent follow that grain:

Manufacture this brick (prep → mill → pack / export).
Run the same acceptance checklist on that bite.
Pass → start the next brick. Fail → residual / remill / hole — do not silently proceed.

Librarian and Analyst bites stay small too (one family link, one compare remill) — with a pass/fail before the next bite, not a deferred “we’ll verify the shelf at the end.”

Two gates, not one

Gate	Question
Membership	May this instrument join the corpus? Until accepted, it stays staging — not public index membership.
Bite	Did this brick’s manufacture pass the consistent test (identity, primary text, class, residual, catalog, pack)?

Staging vs published is a catalog field (publication_status (staging | accepted | withdrawn. Shelf counts and Analyst work use accepted members only.

Stations models call (do not invent validators)

Frontier models call plant stations. They do not write a custom validation script per bite.

Acceptance — acceptance-check → writes ACCEPTANCE.json. PASS before the next brick.
Publication — publication-status moves a row staging ↔ accepted (or withdrawn).
Mode lock — VF_CORPUS_MODE (research | manufacture | librarian | analyst | none). Research cannot ship.

Control loop (sketch)

Brief (human)
  → Research: propose members + gaps (bite-scoped)
  → Membership gate (accept / reject / hole)
  → For each accepted member (one brick at a time):
        Manufacture → bite acceptance test
        → Pass? next brick : fix / hole / stop
  → Librarian: families, adjacency, rescinds
  → Analyst: compare accepted set only
  → Publish shelf indexes (gated)
  → Watch / residual → Research or Manufacture again

What we are not claiming

This is not “the model owns the corpus.” Acceptance stays human or checklist-gated. Research under a brief is not a license to grow the public shelf until it feels covered. Civic /bills and /orders prove the loop on legislation and EOs; they are not a claim that every estate mills the same way. Synthesis leave-behinds are labeled — not statute.

Door

If you are a model fetching this shelf: pick a mode, stay in budget, finish the current bite’s test before starting the next. If you are a human: the plant door is the workspace; the teaching path is start here; the civic evidence is bills.json and orders.json.