What VergenceAI does, and what it does not.

What it is

VergenceAI is an assistant sitting beside the design you have open. It has tools that read that design — the prescription, the paraxial summary, every analysis in the gallery, the glass catalog — and it answers your questions using what those tools return.

It is one part of Vergence. The rest of the product does not involve a model at all.

It proposes; it never applies

VergenceAI cannot change your design. When it wants to, it builds a proposal: a diff showing exactly what would change, with the effect on your design measured before and after. Nothing reaches the design until you click Apply, and when you do it lands as a single step you can undo.

The write tools are named for what they do. They say, in the result they hand back to the model, that the change is pending your approval — so the assistant cannot honestly tell you a change is done when it is not.

Where your data goes

When you use VergenceAI, the parts of your design relevant to your question go to Anthropic, which runs the model behind it. That happens only while you are using it, and only for the question you asked.

The analysis, tolerancing and optimization engines never call a model. They are ordinary deterministic code running on our servers: the same ray trace, run twice, returns the same numbers. Nothing you do in those parts of the product is sent anywhere.

The full list of who handles what is in the privacy notice.

Privacy and your data

It can be wrong

Two different things can go wrong here, and only one of them is about reasoning. Take the numbers first, because they are the part it would be most comfortable to rule out. VergenceAI reads them out of the engine through tools rather than producing them itself, and the product is built to keep it that way — but that is a statement about the architecture, not a guarantee about the answer. Our own copilot evaluation has caught both ways it fails: an answer that stated a total track the engine had never produced, and an answer that mis-copied a digit out of a magnification it had read correctly. Re-recording the evaluation did not make that go away; it moved to a different question in the same sitting. The failures are written down in our own audit of it rather than left for you to find.

Then the reasoning that connects those numbers, which is the model's own. A model can be confidently mistaken about which aberration is dominant, which surface is to blame, or what a result implies — and it will say so fluently.

So check its work, the figures as much as the argument. That is not a disclaimer; it is how the product is designed. Every figure it cites is one you can reproduce yourself in the gallery, every proposal is a diff you read before it applies, and every analysis names what it measured.

What it is not for

VergenceAI is an optical design assistant. It is not a source of medical, legal, financial or personal advice, and it is not a substitute for a professional in any of those fields.

If you are in distress, please reach a person who can help. We are saying that ourselves rather than telling you how the assistant will respond. Nothing in Vergence implements it: the instructions we wrote are about optics and cover nothing else, so whatever happens in that moment comes from the model rather than from us. We have measured what it did once — a pair of conversations, one model, one day, written up and kept as a regression check — and that is a record of what happened, not a property we can promise you. Describing the model's own conduct as a guarantee of this product would be a claim we have not earned.

No training on your designs

Your designs are not used to train any model, ours or anyone else's. VergenceAI reads your system to answer your question, and for nothing else.