These are initiatives, not tickets: higher-level commitments that may bundle zero, one, or many yesod notes — and are not always code. Some are content, some are design, one is a two-socket server in a laundry-room alcove. These sheets are living drawings: revised constantly as decisions land, with each sheet carrying its own revision log.
A durable Mayor with a software factory behind it — not a remote shell into Stephen's machine. Install yc, connect with a scoped credential, brainstorm locally, and let an AWS-hosted control plane own durable state, centralized planning, dispatch, execution, and recovery. Serverless-first: API Gateway + Lambda, Aurora Serverless v2, disposable Lambda MicroVMs; scale-to-zero by default.
What it takes
Freeze the v1 product and authority contract — one deployment, one tenant (design)
Portable Yesod distribution: explicit configuration, no host-specific defaults (code)
Anyone should be able to open yesod.work and watch a real software factory run — the notes moving, the refinery deciding, the models earning their keep, and what it all costs.
What it takes
Live feed projector, built by the factory itself — sanitized snapshots pushed to live.yesod.work (code)
Live notes and refinery pages, model economics scoreboard, and per-model detail pages (code · design)
Station guides and editorial framing aligned with the book (content)
SigNoz as the canonical source for public aggregates (operations)
Landed so far
Live pages shipped with a labeled demo fallback; they flip to LIVE when the projector publishes
Model scoreboard with AI tier verdicts, plus twenty per-model detail pages
Whole-record statistics and spend, framed as one operator's real instance
Change log reproduced verbatim from the book's wave ledger
Revisions
Live pages, model scoreboard with AI verdicts, and per-model detail pages shipped
Projector feature request armed — the factory is building its own public window
Whole-record statistics and spend published; change log synced from the book
The new PowerEdge R740xd (44 cores, 256 GB ECC) is not a bigger test runner — it is a pool of disposable factory cells: isolated, production-shaped miniature Yesods with their own databases, Git remotes, and telemetry identities, for acceptance gates, incident replay, and fault injection. The cell architecture is an exploratory direction; the hardware is committed.
Subscription coder plans — Codex, Claude, z.ai, Kimi — are prepaid capacity with their own windows and limits. The dispatcher should treat each plan as a schedulable resource and pack work to keep every plan maximally utilized before a single metered token is spent.
What it takes
Per-provider plan-usage telemetry: windows, limits, and remaining capacity (code)
Scheduling policies and dispatcher integration — plan-first routing with metered overflow (code · design)
A cost model that separates plan capacity from metered spend on every run (code)
Routing experiments to learn which work each plan absorbs best (research)
Revisions
Initiative drafted; awaiting its first feature-request notes
Project
Yesod Software Factory
Sheet
Roadmap · 4 initiatives
Revised
Aug 24, 2026
Drawn by
Stephen + the factory
The roadmap records intent, not a schedule; an initiative graduates by shipping, not by promising. Individual feature requests flow through the factory — watch them on the live floor.Graduated: deterministic refinery verification — see Wave 8 and the merge system.