The Dark Factory / Field notes
Field notes
What the factory is learning: proof, verification, the models we test and the mistakes we make along the way. 14 notes, newest first.
Latest · 26 Aug 2026Ghillie, and TheReefink
We build in Ada. To use the libraries every program needs — graphics, images, a camera, a UI — Ada needs bindings that, for several of them, didn't exist. So we wrote four and open-sourced them. Three are bindings, not proofs, and we say so; one carries a machine-checked safety layer.
Read the note →-
29 Jul 2026
The duck factory builds provable software.
We stopped work to enter AMD's DevMaster hackathon. In a day we had speech and inference running on Radeon GPUs. The first sentence the machine understood renamed the company.
Read → -
18 Jul 2026
The fourteen-pence run.
We put Moonshot's Kimi and Anthropic's Fable through the proof gate on the same day. One of them cost fourteen pence. One of them took the summit. And the only unanimous failure turned out to be ours.
Read → -
12 Jun 2026
Congratulations. You’re on the floor.
We don't take a model on because it's clever — we take it on because it cleared the gate, and most don't. Congratulations to the models that just earned a seat: Gemini, DeepSeek and Claude.
Read → -
11 Jun 2026
The factory wrote the bug and blamed the workers
Our benchmark accused two language models of a mistake they did not make — the fault was our own tooling. A correction and a public apology to DeepSeek and Google. An honest benchmark owns its errors, especially the ones in its own favour.
Read → -
10 Jun 2026
The new model wrote the old bug
A newer language model joined the factory this week. Unaided, it wrote the exact numerical bug its predecessor wrote three weeks ago. Inside the factory it couldn't. We improved the factory — and our benchmark didn't move, which is the point.
Read → -
5 Jun 2026
The bridges nobody had built
We build in Ada. To use the libraries every program needs — graphics, images, a camera, a UI — Ada needs bindings that, for several of them, didn't exist. So we wrote four and open-sourced them. Three are bindings, not proofs, and we say so; one carries a machine-checked safety layer.
Read → -
30 May 2026
The cobbler’s children
We pointed our verification factory at itself. It failed — not in the proving, but in a small hand-coded tool nobody had thought to verify. On bootstraps, borrowed dependencies, and a changed arithmetic.
Read → -
29 May 2026
A new model arrived this week. Our promise didn’t change.
We put the newest frontier models to a mathematical proof of correctness. They are good — but a model predicts plausible code; it does not prove correct code. That gap is the whole product.
Read → -
24 May 2026
Old code, new code, the numbers.
Show-your-working from today's factory run. Four routines side-by-side, the Fortran in, the SPARK-Ada out, the discharge number. Per-routine version bar chart for all seventeen.
Read → -
24 May 2026
What 96 per cent means when the spec actually says something
Our score went down and we could not be prouder. The first kernel checked under strict rules came in at 96 per cent — and that's the first number we believe.
Read → -
24 May 2026
Where the prover pushes back hardest
Eleven routines under strong postconditions, 91.2% aggregate honest discharge. The interesting part wasn't the headline. It was the spread.
Read → -
20 May 2026
Tens of thousands of lines of safety-critical Fortran, modernised. The audit that halved the number. And the kitten.
Factory bench, week one. USGS NSHMP, LAPACK DGESVD, DL_POLY Classic. The audit. The kitten.
Read → -
20 May 2026
The Senior Service: what the factory borrowed from 480 years of Royal Navy engineering tradition.
What the Dark Factory borrowed from 480 years of Royal Navy engineering tradition: watches, the log, damage control, charts, defence in depth.
Read →