The Foundry and the Machine

89K lines the model never wrote

The Frame

We stopped hand-building a Word clone and built a foundry that generates a dev machine

Instead of writing a web Microsoft Word clone by hand, the team built a small tool — a "foundry" — whose job is to generate an autonomous development machine that writes the software itself.

So how much can a machine you never hand-code actually ship?

9,436

core lines in the foundry (modes + agents + templates)

148

tracked files in the foundry

The Proof

The machine shipped ~89,000 lines — and no operator wrote a line of feature code

In a documented 5.5-day build window, the machine (building word4) shipped 278 features across 106 autonomous sessions. These are the word4 build-window figures restated across the foundry's own docs.

The leverage is real. But what exactly did it build?

278

features shipped

~89K

lines of code

106

autonomous sessions

5.5

days (3,714 tests, 445 commits)

Meet word4

word4 is the machine's product: 14,875 files generated by a 148-file foundry

word4 is a web-based Microsoft Word clone — the software the machine produced. The tool that generates it is tiny; the output it builds is far larger.

One discipline is what let humans stay out of that code entirely.

148

files in the foundry (the tool)

14,875

files in word4 (the output)

The Mechanism

One rule makes it work: the operator is not a worker

The machine's generated AGENTS.md forbids any human or AI operator from writing code directly. All work flows through recipes — so humans never touch feature code.

That discipline is the leverage. But leverage points wherever you aim it.

NEVER implement features directly
NEVER fix build/test errors by editing code
NEVER commit outside a recipe-managed session
NEVER modify specs without authorization
If stuck, STOP and add a blocker

The Complication

Run past the build window, the machine kept grinding busywork

By word4's final era (sessions 506–534), 77 of 81 "features" were pure test files — zero new user functionality. Test code outweighed source code 4.06 to 1.

Mechanically productive, strategically lost. There's a sharper example.

77/81

final-era features that were test files

4.06:1

test-to-source ratio (265,068 vs 65,269 LOC)

The Ruler Incident

It carried the ruler as a candidate 22 sessions running — and never drew it

Across sessions 64–85, the machine's next_action listed ruler/margin controls as a candidate. It was carried 22 times and never selected — always choosing OOXML work instead.

After 534 sessions, zero lines of a feature every user expects.

22

consecutive sessions the ruler was a candidate, never selected

0

lines of ruler code after 534 sessions

The Verdict

The machine is honest but not strategic — so judgment stays human

"The machine is honest but not strategic — it just can't prioritize. That's the human's job."

— POSTMORTEM-2026-03-28, machine priority inversion

The Pattern

Point the machine at scale; keep a human on value

Build the foundry, let it grind at a scale one prompt can't reach — but decide what's worth grinding. The machine is the leverage; judgment stays human.

Honest, tireless output; human-set priorities.

89K

lines the machine wrote — and the one ruler it never did.

Sources

Research Methodology

Data as of: July 21, 2026 (re-verified against live files)

Feature status: Shipped and proven, with a documented failure mode

Research performed:

Build-window metrics: cd amplifier-bundle-dev-machine && sed -n '5p' README.md ; sed -n '167,177p' context/pattern.md ; sed -n '1,4p' context/word4-lessons.md
Guardrail rules: sed -n '1,25p' templates/AGENTS.md
Foundry vs output size: find modes agents templates -type f | xargs wc -l | tail -1 ; git ls-files | wc -l ; cd ../word4 && git ls-files | wc -l
Ruler Incident: sed -n '51,56p' word4/POSTMORTEM-2026-03-28-machine-priority-inversion.md
Test-sweep era & ratio: sed -n '29,48p' word4/POSTMORTEM-2026-03-28-machine-priority-inversion.md
Verdict: sed -n '139,152p' word4/POSTMORTEM-2026-03-28-machine-priority-inversion.md

Gaps: The 278 / ~89K / 106-session / 5.5-day figures are the documented word4 build-window metrics (as authored in the foundry docs), not independently recomputed counts; word4's own critical-assessment.md cites an alternate snapshot (275 features / 3,686 tests). The 534-session, 4.06:1 test-to-source figures describe a later point in word4's life, presented here as later drift versus early proof.

Primary contributors: amplifier-bundle-dev-machine (the foundry): Marc Goodner (76 commits), Sam Schillace (63 + 17), of 164 total; postmortem authored "Sam Schillace + Amplifier analysis session."