Methodology

This site mixes three kinds of content, and the difference matters if you are going to rely on any of it: text reproduced exactly as published, structure derived from that text by deterministic code, and summaries written by a language model. This page says which is which, and what is known about how good the third kind is.

The source

One pinned copy of the USPTO Manual of Patent Examining Procedure, Ninth Edition, Revision 01.2024, fetched from uspto.gov and identified as mpep-e9r01-2024-html. All 736 file-level sections are stored with a SHA-256 hash of the file they came from, so the copy being served can be checked against the copy that was fetched. The site does not track the USPTO site: it is a snapshot, which is what makes a citation to it stable. See the corpus page for the current version and coverage.

USPTO guidance issued after that revision is kept in a separate layer and never merged into the manual text. A section that has been overtaken by later guidance says so in a banner; the text underneath it is still the text as published.

What is reproduced verbatim

MPEP section text, statutory and regulatory text, and examiner form paragraphs are reproduced as published and are not edited, summarised or rewritten. Headings are re-levelled so each page has one top-level heading, but the words inside them are untouched. Statutory text comes from the manual’s own appendices — 330 regulations from Appendix R and 100 statutes from Appendix L, both at Revision 01.2024 — so the statute you read here is the statute the examiner’s manual reproduces, at that revision, not the current U.S. Code.

What is derived deterministically

Splitting the corpus into sections and subsections, the anchor for every subsection heading, cross-reference links between sections, links from citations into the statute and case pages, the chapter groupings, and the full-text search index are all produced by code from the source files. Running it again on the same input gives the same output. No model is involved, and a mistake here is a bug that can be reproduced and fixed.

What a language model produced

Two things. First, the structured annotation layer: for 425 of the 736 sections, the citations in the text have been turned into typed relations — which statute, which case, for what proposition — with quotes carried back to the source. Sections that have not been through that pass are still fully readable; they are just not machine-structured, and the section page says which it is.

Second, the case holding summaries. Every one of the 1,729 cases the manual cites has a one-sentence summary, 1,729 in total, and every one was written by a language model. They were generated by claude-fable-5 (batch agent pass, 2026-07-02); claude-opus-5 (phase8 chapter 1800, 2026-09-02, +1 holding).

An important limitation in how they were made: the summaries are grounded in how the MPEP uses each case — the surrounding prose, the annotator’s proposition, and the manual’s own parentheticals — rather than in an independent reading of the opinion. So a summary tells you what the case stands for as the manual deploys it. That is usually what you want when working from the MPEP, and it is not the same thing as the holding of the case.

What the confidence levels mean

Each summary carries high, medium or low: 1,567 high, 132 medium, 30 low. Being straight about this: the level is the generating model’s own self-assessment, recorded at the time it wrote the summary. It is not calibrated against a measured error rate, and no accuracy study of these summaries has been published. Treat it as a rough signal of where to look harder, not as a probability.

Human review

No holding summary has been reviewed by a person yet. Review is being set up, and when a summary has been checked its case page will name the reviewer and the date. Until a page says that, assume it has not been.

Because of that, and because the summaries describe the manual’s use of a case rather than the opinion, read the opinion before relying on a holding in a filing. Full opinion text is available on site for 1,400 of 1,729 cited cases, from Caselaw Access Project (Harvard Library Innovation Lab) (1,347), CourtListener (Free Law Project) (52) and U.S. Reports vol. 573 (supremecourt.gov bound volume), text via datalab Marker (1).

Versioning

The corpus is identified by a version string that names the edition and revision it was built from. Section text changes only when a new MPEP revision is ingested, which produces a new version string rather than editing the existing one. The annotation layer and the holding summaries can be regenerated between corpus versions; the review status above is stored separately from both, so recording a review does not require rebuilding either.

Corrections

Errors in this site’s parsing, linking or summaries are worth reporting and get fixed. How to report one.