AI Terms

In one place: the CAD tooling our mechanical team works in. An engineer describes a part in words and the engine answers with a parametric script that builds it. That script is a starting point an engineer then edits, not a finished component.

Separately, our machine-learning work trains models on imagery data sets we own, on rented GPUs. Neither the training data nor the models come from client work.

The Subprocessors page names the companies behind both, and the Privacy Policy covers what leaves our infrastructure when the engine runs.

Work we hand over may have been drafted with model assistance somewhere in its history. That changes nothing about who owns it: the Terms of Service govern that, and a Deliverable becomes yours once the engagement it belongs to has been paid for.

• An engineer is answerable for it either way. Nothing reaches you because a model produced it. It reaches you because a person reviewed it and put their name on the work.
• Ask and we will tell you. If you want to know whether a model was used on a particular piece of your project, ask in the thread. Some clients need to know for their own compliance, and it is a fair question.
• Tell us if the answer has to be no. Where your policy or your customer forbids model-assisted work, say so during scoping. We will either agree to it in the proposal or decline the work.

Model output can resemble material somebody else holds rights in, and a purely machine-generated result may not attract copyright at all in some countries. Where that distinction matters to what you are building, raise it before we start rather than after handover.

There is no chatbot, no assistant and no automated agent on this site. A message you send through an inquiry, a support ticket or the contact form is read and answered by a person, and no model decides anything about you or your project. Anyone telling you otherwise while claiming to be SpaltX is not.

The disclaimer in the Terms of Service covers everything we deliver. Two consequences are specific enough to model output to be worth stating here rather than leaving you to derive them.

• Check the numbers. A dimension, a tolerance or a material property that came out of a model is not a fact until somebody has verified it against the source it claims.
• The same prompt can answer differently. Nothing about model output is reproducible in the way a machined feature is, so it is never the record of what was agreed. The proposal is.

Where a third-party model provider gives us protection against a claim that its output infringes somebody’s rights, that protection has limits set by them, not by us. It typically falls away where the output has been modified, combined with other material, or driven by an input the customer supplied. Producing engineering work does all three.

So we do not offer you more than we hold. Your remedy for anything model-related is the one in the Terms of Service, and the indemnity running the other way is unchanged: what you send us has to be yours to send.

Two commitments, and they are not the same one said twice:

• We train nothing on your work. Not your files, not your correspondence, not the geometry an engineer built for you. Our own models are trained on imagery data sets we own.
• We do not permit a provider to either. The model provider behind the design engine is engaged on business terms that do not allow training on what we send. Those terms are theirs to publish, so read them at the source rather than taking our summary for it, and ask us and we will tell you which provider and which terms.

If your material is sensitive enough that a prompt leaving our infrastructure is itself the problem, say so at scoping. The deterministic engine, forge-de-1, runs entirely on our own machines and sends nothing anywhere.

This page moves when what we do moves. A new provider appears on the Subprocessors page in the change that connects it, a material change here gets the notice described in the Terms of Service, and questions go to support@spaltx.com.