Accessibility work involves documents that are often confidential — payroll, medical billing, financial statements, case files. Customers reasonably want to know whether those documents are being fed to a model. This page answers that plainly, and we will keep it accurate as the service changes.

The rendering pipeline contains no AI

Converting HTML to a tagged PDF is a deterministic problem, and we solve it deterministically. Layout and text shaping are done by WeasyPrint and Pango. Structure and metadata are written with pikepdf. Conformance checking walks the PDF structure tree and evaluates the machine-verifiable checkpoints of the Matterhorn Protocol directly.

The same input produces the same output. Nothing in that path makes a probabilistic guess about your content, which is a property worth having when the output is a compliance artefact.

What that rules out

Some tools in this space generate alternative text for images, or infer a reading order the source markup does not express. Those features need a model, and a model needs to see the document.

Wellform does not do this today. The consequence is honest and worth stating: if your HTML has no alt text, we will report the failure rather than invent a description. A generated caption can be confidently wrong in ways a reader cannot detect, and for a document you are signing your name to, a flagged gap is more useful than a plausible guess.

If that changes

We may later offer AI-assisted features — suggested alternative text is the obvious candidate. If we do, we commit that any such feature will be:

• Opt-in. Off by default, requiring an explicit flag on the request. It will never be applied to documents from customers who have not asked for it.
• Disclosed. This page will name the feature, the provider processing the content, and what they receive, before the feature ships.
• Labelled in the output. Machine-generated content will be identifiable as such in the conformance report, so it is never mistaken for something a person wrote and checked.
• Never training data. Content sent to any provider will be under terms that prohibit training on it.

AI in how we build Wellform

We use AI coding tools to help write and review the software, as most engineering teams now do. This does not involve customer data: the code is written against test fixtures and public specifications, not against documents anyone has sent us. Changes are tested and reviewed before release regardless of how they were drafted.

Using Wellform inside an AI system

Wellform is a plain HTTP API and works well as a tool call inside an agent or generation pipeline. If you are producing documents that way, note that our conformance report describes the file, not the truthfulness of what your system wrote into it. A perfectly tagged PDF full of fabricated figures is still a perfectly tagged PDF. Accessibility and accuracy are different problems.

Contact

Questions about any of this go to support@wellform.dev.