The factory,
with the lights on.

The idea

Most claims about agentic software factories are marketing. The strongest counter-argument is a window: open yesod.work and watch a real factory run — the notes moving through their lifecycle, the refinery deciding what lands, the models earning (or losing) their tiers, and the money it all costs. Not a demo reel; the actual instance, with its failures on the record.

The factory’s own telemetry, published with the same discipline the factory applies to code.

Architecture

Nothing public ever connects into the factory’s network. A projector — itself built by the factory as a feature request — reads the internal live-viz API, projects records field by field against published contracts, and pushes sanitized snapshots outward to object storage. The public pages poll those snapshots and flip from OFFLINE to LIVE on their own.

   factory (private network)            edge                    browser
┌─────────────────────────┐    ┌──────────────────┐    ┌──────────────────┐
│ live-viz API            │    │ R2 object store  │    │ /live/notes      │
│      │                  │    │ live.yesod.work  │    │ /live/refineries │
│      ▼                  │    │                  │    │                  │
│ projector               │───▶│ notes.json       │◀───│ poll on the      │
│  · allowlist 3 projects │PUT │ refinery.json    │GET │ snapshot's own   │
│  · field-by-field       │    │ health.json      │    │ refresh hint     │
│  · fail-closed scan     │    └──────────────────┘    └──────────────────┘
└─────────────────────────┘

The privacy contract is structural, not aspirational: only allowlisted projects; only allowlisted fields; any record that trips the credential/hostname scan is withheld entirely and counted — the pages display the withheld number rather than pretending completeness.

What has shipped

• Live pages — /live/notes and /live/refineries, with LIVE / STALE / OFFLINE states and a labeled demo fallback while the projector is in the factory’s queue.
• Model economics — the scoreboard of every model ever run (2,155 runs across 20 models), with AI tier verdicts and twenty per-model detail pages: duration distributions, tokens, cost per run, failure habits, idiosyncrasies, lessons.
• The whole record — run statistics (outcome bars, activity heatmap, peak concurrency) and spend (metered $1,235 plus a labeled estimate to ~$2,655 all-time), each framed honestly as one operator’s personal instance.
• The change log — all ten development waves, reproduced verbatim from the book’s wave ledger.

What remains

• The projector itself, moving through the factory as ys-yes-c1ns — when its first snapshot lands, the live pages light up with no site deploy.
• SigNoz as the canonical source for the public aggregates, replacing the per-page capture scripts.
• Phase 2, if wanted: true realtime push (worker-relayed SSE/WebSocket) instead of snapshot polling — polling with row-flash animations is the accepted experience until the demand is real.

Why it matters

The explorer is the factory’s honesty mechanism. Publishing the record — including the 43%-success model, the 866 cost-blind runs, and the withheld counts — makes the whole story credible in a way no landing page can.