Models in the factory
Five seats in this factory are occupied by a model. Everything between them — the merge controller, the gate, the promotion transaction — is deliberately deterministic. This page names each seat, the route it runs on today, and how the factory classifies the models it employs.
Five seats
A seat is a contract: what goes in, what must come out, what gate judges it. The occupant is a route — harness, provider, model, payment path — and routes get swapped when the ledger says a cheaper or better one exists. Several of these seats have changed occupants without the seat's contract moving at all.
coder plan
Orchestrates the factory: plans campaigns, files and dispatches work, watches the line, intervenes when a contract parks something, and writes the autoblog when something interesting happens.
Route today: Claude (frontier tier) in the Claude Code harness, running on a subscription coder plan.
coder plan first, metered fallback
Turns a captured note into planned work: a bounded Bead tree with dependencies, merge boundaries, and acceptance criteria. Ambiguity tolerance and decomposition quality matter more here than speed.
Route today: A frontier model on the Codex coder plan, invoked through the registered Codex backend. When plan capacity is exhausted, planning falls back to a pay-per-token lane rather than stalling the queue.
mixed — metered volume, plan-covered heavy lifting
The runner fleet executes planned work in isolated worktrees. Each lane is one registered route — harness, provider, model — and the dispatcher matches a note’s complexity class to a lane’s tier: easy work goes to inexpensive lanes, advanced work to expensive ones.
Route today: Many lanes at once, deliberately: pay-per-token lanes on Fireworks (DeepSeek, GLM, Kimi families) carry the volume; coder-plan lanes (Claude Code, Codex) carry the hard cases. The observed mix below is read from the public run ledger.
metered
Answers questions over the tool registry — yesod ask and its distill pipeline — returning cited answers from accumulated notes rather than guesses.
Route today: Kimi K2.7 Code on Fireworks, pay-per-token. Costs pennies; runs constantly.
metered — cents per month
Screens every note body before it is published to the public live feed, proposing redaction spans for operator review. It proposes; it never edits. A deterministic credential gate runs before and after it, so the model is never the last line of defense.
Route today: DeepSeek v4 Flash on Fireworks — pinned in the projector’s source, not configurable by environment, budgeted to four new screening calls per publish cycle. Verdicts are cached by content hash, so each note revision is screened exactly once.
Observed, not asserted
Notes carry the model that carried them. This table counts the public feed's notes by model — recomputed in your browser from the live feed when it is reachable, otherwise shown as of 2026-08-19.
| Model | Notes carried |
|---|---|
| deepseek-v4-pro | 290 |
| glm-5p2 | 192 |
| claude-opus-4-8 | 156 |
| kimi-k2p7-code | 118 |
| gpt-5.6-sol | 49 |
| claude-fable-5 | 39 |
| kimi-k2p6 | 29 |
| kimi-k2.6 | 22 |
Snapshot from the public run ledger, 2026-08-19.
Deliberately model-free
The standing rule: models make the work; machinery decides what counts. No model verdict admits code to main, promotes a release, or closes a note. The dispatch queue and claim mechanics, the merge controller, the deterministic gate, and the release-promotion transaction contain no model call. Each advances only on named evidence — a result commit, a green gate snapshot, a completion marker — and fails closed when the evidence is missing. A model can be wrong all day inside its seat; the guarantees between the seats are not allowed to be.
How models are classified
The factory does not keep a "best model" opinion. It keeps a routing input, a measured outcome, and a payment path — and lets the scoreboard argue.
The scoreboard
Every model's dispatch tier and earned S–D verdict, from terminal attempts.Open model economics →The roadmap
Plan-first routing: spend prepaid coder-plan capacity before metering.Coder-plan scheduling initiative →