Considering an AI tool for AEC? Ask about its Deterministic Safety Harness

Considering an AI tool for AEC? Ask about its Deterministic Safety Harness.

Architecture, engineering and construction has spent a year learning that general purpose AI is remarkable at almost everything except being precisely right. That lesson has produced a pitch. Generative AI is probabilistic, our industry is not, so the AI output is passed through a deterministic engine that checks it against the rules before anyone trusts it. That story was told from a dozen stages at this year's Autodesk events, and it is being told in design offices, estimating rooms and site huts right now.

It is a good story. It is also where the conversation stops, and it stops because nobody on those stages had a solid term for the thing being described. Without an agreed term, a verification layer is a slide. It cannot be written into a contract, compared between vendors or asked for by name.

‍

The term exists. It is Deterministic Safety Harness.

‍

And it’s already in use in AEC. At ConnectionAI, we’ve been thinking about our Deterministic Safety Harness (DSH) since inception, because the gap between probabilistic and deterministic is widest at LOD400, where precision is the entire problem. It is why the industry has been slow to trust generative tools, and it is the problem AI built for AEC has to solve.

A DSH is a verification layer that sits between what an AI generates and what a professional is asked to sign, price or build. It applies fixed checks, traceable to the governing code or standard, to every generated element, and it applies them the same way every time. It does not learn and it does not guess. It records what was checked, against what, and with what result, so when an output reaches a human it carries a paper trail rather than a probability.

The shape of the harness changes by discipline but the principle does not. Every element the AI generates is checked against the rule that governs it, using the method a qualified professional in that discipline would use, before anyone is asked to rely on it. In every case the liability for what gets built does not sit with the software vendor, it sits with the licensed professional and the contractor who put their name to it. Any AI that produces deliverables without a harness in front of it is asking the industry to be the harness.

Put that way, the DSH sounds like quality control, and most of the industry will treat it as quality control. That undersells what should actually be happening.

‍

The DSH is AEC’s version of AI alignment.

‍

The frontier AI labs treat alignment as the central problem of their field. OpenAI publicly committed a fifth of its computing power to it. Every major lab now publishes a safety framework describing what its models may and may not do. Their alignment problem is defined by human values, which are contested and hard to write down. AEC's alignment problem is defined by physics, by code and by contract, which are neither. The right answer is knowable. The tolerance for being wrong is zero, because the outputs become buildings people occupy. An alignment problem with a known answer and no room for error deserves at least the seriousness the labs bring to theirs. So far it has received a fraction of it.

Alignment is also much more than a set of rules. A rulebook tells a model what it may not do. Alignment shapes what a model wants to do, and the labs learned early that the difference matters. Filters bolted onto a finished model are brittle. Real alignment happens in training, in the data the model sees and the signal it is optimized toward, and external checks exist to confirm that training worked rather than to substitute for it.

That principle exposes the limit of most tools currently on the market. They wrap a harness around a general purpose model that was not trained for our industry. When two frontier models were asked to connect a real steel structure, they located the joints with near perfect accuracy and then produced connections that were unusable in every case, with plates buried in webs and bolts rendered as cubes. Those models will improve, and one day they may well close the gap. But a harness in front of a model the vendor does not control can only catch failures. It cannot reduce them, because nothing the harness learns ever reaches the underlying model.

A harness whose verdicts flow back into training does something different. Every output it passes and every one it rejects becomes evidence the next model learns from. Over successive generations the model stops being caught by the checks and starts learning them. The harness never goes away, because trust in this industry is earned through verification and not through promises, but the gap between what the model generates and what a professional can stand behind closes with every cycle.

Embedded alignment for AEC looks like precision and trust in real training data rather than whatever the open internet holds about buildings. It looks like academic rigor in evaluation, measuring whether an output is buildable, compliant and priceable, not whether it looks plausible. And it looks like a harness that is created, deployed and then used to train the AI model that follows it. Not a filter bolted onto someone else’s model to make the output more palatable. Alignment is capability, and structural steel is one domain where the AI alignment loop is already being built. It should be every domain.

This is the way AEC builds trust in the future tools that will build the world. So the next time an AI tool is being considered for your business, ask one question before the demo starts. Does it have a Deterministic Safety Harness and do its results influence the AI model?

‍