The Matterhorn Protocol: 136 ways a PDF can fail PDF/UA
PDF/UA-1 checkers test the conditions of the Matterhorn Protocol: 31 checkpoints, 136 failure conditions, and the split between machine checks and human judgement.
, updated , a 3-minute read
If you have run a PDF through PAC, veraPDF or our online checker, the checks behind the results go back to one document: the Matterhorn Protocol.
PDF/UA-1 (ISO 14289-1) is the standard. The Matterhorn Protocol, published by the PDF Association’s PDF/UA working group, is its testing model: a list of every way a PDF can fail PDF/UA-1.
What is in it
The protocol organizes PDF/UA-1 into 31 checkpoints, groups such as “real content tagged”, “headings”, “tables”, “alternative descriptions”, “fonts” and “metadata”. Inside them are 136 numbered failure conditions, each a concrete statement of what is wrong. A few real examples, paraphrased:
- Real content is neither tagged nor marked as an artifact (checkpoint 01).
- Heading levels skip: an H3 follows an H1 with no H2 between them.
- A
Figuretag has no alternative description and no replacement text. - A table cell cannot be associated with any header cell.
- A font’s characters cannot be mapped back to Unicode, so extracted text is wrong.
- The document does not declare the language it is written in.
Each condition carries a number, such as 01-006 or 14-003. PAC reports by these checkpoints. veraPDF, and our checker built on it, report the clause and test of ISO 14289-1 instead (“clause 7.18.1, test 2”); the two numberings describe the same requirements.
Machine-testable and human: 87 and 47
Every failure condition is classed as machine-testable or needing human judgement. In version 1.1, 87 are machine-testable, 47 need a person, and 2 have no specific test (older tallies say 89 and 45, from before two conditions moved from one class to the other). The human ones are about meaning:
| A machine checks | A person decides |
|---|---|
| that an image has alt text | whether the alt text describes the image |
| that heading levels do not skip | whether the headings describe the document’s structure |
| that content sits in a reading order | whether that order makes sense |
So “0 errors” from any checker, ours included, covers the machine part only; the protocol names what needs a person.
How the tools map to it
- PAC (free, for Windows) implements the machine-checkable failure conditions and reports by checkpoint; our page about it compares it with the online checker. Its Screen Reader Preview and structure viewer help a person work through the other 47 conditions.
- veraPDF (open source) implements PDF/UA-1 as a validation profile: machine conditions only, written as rules over the PDF’s objects. It is the reference validator, and the one behind our checker.
- Our online checker runs veraPDF’s PDF/UA-1 profile on our server and shows each failed rule with the elements it found on the page, what passed, and a link to the rule; the common failures are explained in plain words. It does not try to automate the human part; it lists what is left for you to review.
Matterhorn and PDF/UA-2
PDF/UA-1 (ISO 14289-1, first published in 2012, second edition 2014) is written against PDF 1.7. Its successor, PDF/UA-2 (ISO 14289-2, 2024), is written against PDF 2.0 and comes with the Well-Tagged PDF (WTPDF) 1.0 specification; veraPDF has validation profiles for both. The Matterhorn Protocol covers PDF/UA-1. PDF/UA-2 has its own post.
What to do with it
Run every document through a machine check: veraPDF, PAC or the online checker. Give the 47 human conditions a person’s time: alt text, reading order, the use of headings. A tool that writes tags while it makes the PDF can meet the machine conditions at the source; ReflowPDF’s exports pass veraPDF’s PDF/UA-1 profile by default.
The Matterhorn Protocol is a free download from the PDF Association, as a tagged PDF.
Questions
- What is the Matterhorn Protocol?
- A testing model for PDF/UA-1 published by the PDF Association’s PDF/UA Technical Working Group. It enumerates every way a PDF can fail ISO 14289-1: 31 checkpoints containing 136 numbered failure conditions, each classified as machine-testable or requiring human judgement. PAC reports by its checkpoints.
- How many Matterhorn checks can be automated?
- In version 1.1, 87 failure conditions are machine-testable, 47 need human judgement, and 2 have no specific test. Older tallies say 89 and 45, from before two conditions were moved from machine to human. A machine can check about two thirds; whether alt text is meaningful, reading order logical and headings used for structure is the author’s part.
- Where do I download the Matterhorn Protocol?
- Free from the PDF Association at pdfa.org/resource/the-matterhorn-protocol/; the document itself is a tagged, PDF/UA-conformant PDF. The current version is 1.1 (2021).
- Is there a Matterhorn Protocol for PDF/UA-2?
- The Matterhorn Protocol is written for PDF/UA-1. PDF/UA-2 (ISO 14289-2, 2024) comes with the Well-Tagged PDF (WTPDF) 1.0 specification, and veraPDF has validation profiles for both.
- How do I test my PDF against the machine-checkable part?
- Run a validator that implements the machine conditions: veraPDF (open source), PAC (free, for Windows), or our online checker, which runs veraPDF’s PDF/UA-1 profile on our server, shows each failure with its place on the page and explains the common ones in plain words.
Check a PDF of your own against PDF/UA-1 with veraPDF, free, in the browser.
More from the blog
PDF/UA-2 (ISO 14289-2): what changes, and how we verify it
PDF/UA-2 is written on PDF 2.0. Five changes from PDF/UA-1, how to check each, and what veraPDF does not check., a 9-minute readTagged and archival PDFs from ReflowPDF: the files, and how to check them
Exports that validate as PDF/UA-1, PDF/UA-2 and PDF/A-3a in veraPDF, each with its report. What the tags are, what the archive adds, and what no validator can judge for you., a 5-minute readHow we measure CSS conformance on the web-platform-tests
We changed the unit behind our conformance headline: a reftest counts as one test, a scripted file counts as its subtests. Both scales, same dataset., a 5-minute read