clockwright clockwright

01 · nothing to maintain

No specialists. No project.
Nothing for you to maintain.

The engine is the part you never build. It comes already built and already checked; your plant is a document the AI agent writes and your engineers can correct.

the usual way clockwright

How work piles up

A specialist builds it. From then on it is software your team pays to keep alive.

Comes with the product, already built and already checked. Nobody on your side touches it.

Your factory

Written in code, by whoever knows the tool.

A document the AI agent writes and your engineers can read and correct.

Is the model right?

Usually nothing checks, and a wrong rule does not crash: it shifts the answer and looks like a result.

The engine refuses a model that does not hold together, then reruns last year to see whether it reproduces what really happened.

02 · last year first

Before it answers anything,
it has to reproduce your last year.

A gap between the model and what really happened is the finding, and the model is corrected until it closes. Ask for a stretch the model was never shown: a model fitted to one period will reproduce that period, which is not prediction.

03 · exact mathematics

Checked against exact mathematics
49 times over.

Erlang C, Pollaczek–Khinchine, Cobham, Burke, BCMP, mean value analysis, Gaver, Palm and Naor among them: forty-nine checks across ten families, each against an answer that mathematics settles rather than another simulation.

All 49, one per test →

04 · a published plant

The hardest published plant,
and how close we are.

SMT2020 is a 1,043-tool semiconductor plant with a published reference run. Our cycle time is within 1.4% of it on one configuration and 0.8% on the other, starting exactly the lots it does. What closed the last gap was not a rule but which of several identical tools takes a lot: a detail no closed-form test can see, found by asking the reference's own reports where its tools spend their time.

What still misses is changeover time on the few tool groups that carry one — and we say so on the page that measures it.

The measurement, and what still misses →

05 · speed

Fast enough to try
1,000 versions overnight.

The 1,043-tool plant above simulates at 12.6 seconds per simulated year on one core of an AMD Ryzen 7 7840U, and 1,000 what-ifs of it, each run eight times, take about nine hours on the same laptop; a small plant's take seconds.

200,000 customers through an M/M/3 queue in 0.076 seconds, single-threaded: 24× faster than the quickest other engine we could run. Every engine had to produce the correct answer before its time counted.

The measurements →

keep reading

How it works →The six steps, the model document, and the three ways to run it Why not just… →A spreadsheet, your scheduler, a digital twin, a consultancy