Computation somewhere else

The five questions

They apply to a cloud solver, a shared model environment, a monitoring platform and a digital twin alike.

Where is the data? Physically, in which jurisdiction, under whose law. This has contractual and sometimes regulatory consequences, and the answer is often not obvious from the interface.

Who can reach it? Not who is supposed to — who can. Every collaborator, every subcontractor, every administrator at the vendor, and everyone who still has an account from a project that finished two years ago.

What happens if access is lost? The vendor changes terms, the subscription lapses, the connection fails on the day of a deadline, the company is acquired. Is the model recoverable in a form another tool can read?

What happens if it is wrong? A shared current model means a wrong assumption propagates instantly to everyone downstream. The single-version benefit and the single-point-of-failure risk are the same property.

Who is liable? For the analysis, for the data, for the availability. Usually less clearly allocated than for a drawing.

None of these is answered by the technology. All five are engineering and commercial questions that need answering before the workflow is relied on.

The link nobody sets up

Every link above is a considered arrangement, negotiated and contracted. There is one that usually is not: an individual pastes model data, a calculation or client correspondence into a hosted AI assistant to summarise, check or draft from.

It is the same five questions, and they have the same answers — the data is now on someone else's infrastructure, in an unknown jurisdiction, reachable by people you cannot name, possibly retained, possibly used to train a model. What is different is that it happened in a few seconds, without a contract, and it cannot be undone. Deleting the conversation does not recall what was sent.

So: client data does not go to an external service without the client's explicit consent, obtained first. Not a judgement afterwards that the content was probably harmless — consent beforehand, because the decision is not the engineer's to make on the client's behalf. Where a practice has an approved tool with terms that cover this, use that. Where it does not, the answer is that the question has not been settled yet.

Versioning

On a local file, version control is a discipline. On a shared model it is a safety requirement, because the failure mode changes.

Locally, an unversioned model means you cannot go back. On a shared model, it means you cannot establish what the analysis was run against. If the geometry changed on Tuesday and the analysis ran on Monday, the results describe a structure that no longer exists — and with everyone working from the same current model, there is nothing to compare against.

Which is why Module 13's model record includes the software version and the date. On a shared model it needs the model version too, and 'the current model' is not a version.

Monitoring and digital twins

A structure with sensors produces data. The term 'digital twin' describes a model kept in correspondence with the real structure using it.

Worth being clear about what makes one useful rather than decorative:

A visualisation shows sensor readings on a model of the structure. Genuinely useful for situational awareness and not a twin.

A twin is updated by the measurements: when the measured response differs from the predicted one, the model changes. That requires deciding in advance what may be updated and by how much — because a model that adjusts freely to match its measurements will match anything, and has stopped being a prediction.

The engineering question is the same one Module 13 asked about validation: what would the measurements have to show for the model to be wrong? If nothing would, the twin is not telling you anything.

Stable principles and dated claims

Everything above is a principle: it will be true of whatever platforms exist in ten years, because it follows from data being somewhere and someone being liable.

What is not in this course is any claim about what a particular platform can currently do, what it costs, or how fast its solvers are. Those figures are out of date before they are published, and a course that carried them would be teaching a snapshot.

The course's currency register tracks 16 time-sensitive claims. Eight are stable principles that do not date. Seven require verification against a current source and are marked as such wherever they appear — including cloud capability, quantum computing, BIM standard editions, data-protection obligations and the carbon factors in Module 20. One is reported as a 2022-era position rather than as current practice.

None is claimed as verified. That is an honest state for a course to be in, and the register says so on its own page rather than leaving a reader to assume.