Computation applied to the built environment.

Architecture, engineering and construction is a field where decisions cost a lot to make, take a long time to fix, and are sometimes dangerous to get wrong. That is exactly why machine learning can help here. It is also why nobody can accept a system that gives an answer without explaining it. Being useful and being accountable do not pull against each other. They are the same problem, and we build as if they were.

In the Theaetetus, Plato asks a question that has never gone away: what is the difference between knowing something and simply believing it and happening to be right? The answer he tests is that you can give an account of it. It did not satisfy him, and people have argued about it ever since. It is still the standard an engineer is held to. A model that returns the right clearance but cannot say where the number came from has produced a true belief. It has not produced knowledge. Knowledge is what somebody has to sign.

The second commitment is about responsibility rather than knowledge, and it is older still. Responsibility has always needed someone who knew what they were doing and could have chosen otherwise. A statistical model is neither of those things, so responsibility cannot be handed to it. It stays where it already was: with the person whose name goes on the drawing. Good software reduces how much that person has to check and shows them where to look. It does not replace their judgement. Software sold as if it did is selling something it cannot deliver.

Everything else we hold to comes out of those two, and none of it is optional. When a system cannot establish something, it has to say so. Admitting ignorance is where this tradition starts, and a confident guess costs far more to discover later than a blank does. Missing information is reported as missing, never quietly filled in. When software uses a model, it should say which one, what it was trained on, and what it is bad at. When it uses none, it should not borrow the word. Much of the credibility problem that artificial intelligence has in this industry was created by software that claimed more than it did, and the repair for that is not better marketing.

A client’s project data belongs to the client. We do not train on it, and we do not treat what passes through our software as ours to keep. The people who use our tools are not a step towards our next model. An account you cannot give twice is not an account at all, so we record versions, rules and inputs. The same question asked a year later returns the same answer, and a disagreement can be settled by looking instead of arguing.

None of this is a wish. It is the set of terms we are willing to be judged by, and the set we would apply to anyone selling this industry software that reasons.