The factory is just another kind of abstraction.
Humans gained control over nature by way of abstraction and then making tools. When you find you can generate electricity with a steam engine, now you have the practical ability to power factories and, eventually, robotics and computers. This is what happened the last time we made a similar level of discovery: it produced increased automation and accuracy.1
If software factories now exist, it means code becomes commoditized, mass produced. It's going to scale dramatically.
Think of each LLM like a human worker in a car factory under Henry Ford. Individual humans make a lot of mistakes. Ford and his engineers organized production around interchangeable parts, specialized tasks, and a moving assembly line. When organized the right way, the Model T could be mass produced at a much greater scale. This essay is inspired by Henry Ford's discoveries about organizing production.2
I'm also inspired by Six Sigma's principles. Measure the defects, reduce variation, and improve the process. Credit goes to Bill Smith, Bob Galvin, and Mikel Harry for their work on Six Sigma at Motorola, and to Walter Shewhart and W. Edwards Deming for the earlier work on statistical quality control and improving production.3
On October 30, 2026, Fortune reported that Cisco's Jeetu Patel had given 27-year-old Padmanabhan Krishnamurthy unlimited AI tokens to refactor legacy code. The plan was for a two-person team, with a four-week demonstration and a goal of refactoring 98% of Cisco's legacy code within 24 months. That is why I think we are in that time now.4
This is where LLMs are now. They make a lot of mistakes. But if organized right, with the right deterministic checkers, the final output of the factory can be software (or hardware) with a very low defect rate.
That is the idea behind ZenoFCIS. I am building a software factory where you describe the behavior you want, LLMs help produce the implementation, and deterministic checkers check the result against the specified behavior.5
The code that fails a check is like a part that fails inspection. It goes back to be fixed. The checks still have to check the right things. A factory can produce the wrong thing very accurately if its instructions are wrong. High assurance needs evidence for the properties being claimed.
And that defect rate being low is the high assurance part. Once you get that defect rate down, then it's all about increasing performance at low defect rates.
A software factory. Every competent developer in the next few years will be using one.
The factory analogy, in four parts
Carry these ideas forward
Mental toolkit
Four handles for the argument. Open a card to recover its definition, limits, and evidence.
The factory
Workers → tools → checked output
The factory is just another kind of abstraction.
Definition, limits, evidence
Meaning. Organize LLMs, tools, and checks into a repeatable process for producing software.
Boundary. The Ford analogy explains the organization of production. It does not establish a software defect rate.
Deterministic checkers
Produce → check → repair
The checks still have to check the right things.
Definition, limits, evidence
Meaning. LLMs help produce implementations. Checkers evaluate whether they satisfy the specified behavior.
Boundary. A verdict covers the specified properties within the stated model and assumptions. The human still has to judge whether the specification captures the intended behavior.
Six Sigma
Measure → reduce variation → improve
Measure the defects, reduce variation, and improve the process.
Definition, limits, evidence
Meaning. The quality of the production process can be measured and improved along with the output.
Boundary. This is an influence on the approach. A measured sigma level would require defined defects, opportunities, and observations from the actual process.
Increasing performance
Low defect rates → increasing performance
Once you get that defect rate down, then it's all about increasing performance at low defect rates.
Definition, limits, evidence
Meaning. Improve production speed, cost, and software performance while preserving the required checks.
Boundary. The Cisco example is a reported plan. Actual throughput and defect rates require their own measurements.
The symbols are shorthand for the scoped definitions above. Machine-readable cards.