Verbatim claim language is a property you enforce, not a prompt you hope survives

Ask any AI model to quote a patent claim exactly, and it usually will. Usually is the problem.

A rejection chart is only useful if every limitation in it is the actual claim language — not a fluent paraphrase, not a helpful simplification, not the model’s memory of what claims like this tend to say. A paraphrased limitation in a chart is worse than a missing one: it looks right, it reads right, and the argument built on it quietly stops being about the claim that will actually be examined.

Prompts ask. Code checks.

Our office-action workflow drafts its rejection chart with a frontier model at high reasoning effort, and that model is instructed to quote limitations verbatim. Then we don’t trust it.

Every chart row passes through a deterministic fidelity check that compares the quoted limitation against the pending claim text itself. Not an AI reviewer — a mechanical comparison. If the row paraphrases, it fails, and the drafting step re-runs with the failure as feedback. A second AI validator reviews the chart’s substance, but the verbatim property specifically is enforced by code, because “another model agreed” is not the same as “the string matches.”

The interesting part is what the check must accept

A naive exact-match check would reject valid quotes constantly. Real claim text contains formatting that survives differently through document extraction: inline equations rendered as images, an ellipsis written as k=1, 2,..., or N, whitespace that shifts across formats. The fidelity gate normalizes both sides so legitimately odd claim text passes while paraphrase still fails. Getting that boundary right — strict about words, tolerant about rendering — is where most of the engineering lives.

Why this matters beyond charts

The same principle runs through the whole platform: the properties that make AI output safe to rely on should be enforced by the engine, not requested in the prompt. Firms can customize the instructions our workflow agents run on — tone, argument ordering, house conventions. What they can’t do, and what nobody can do, is edit away the checks. Customize the judgment, not the guardrails.