Loading…
Loading…
Dataset 20260827-2142-2f1b575 · measured 2026-08-27 21:42 UTC
Web Platform Tests · 23,994 reftests · 81 suites
ReflowPDF renders HTML to PDF without a browser. This is how the engine scores against the same reftest corpus the browsers are judged by, suite by suite and test by test, failures included. Every figure links to the run it came from.
Conformance, this run
89.96%
17,590 of the 19,554 tests this engine can be judged on render identically to their reference. Counting the entire corpus, including every test we cannot run, it is 73.31%.
Measured 27 Aug 2026, 21:42 UTC · dataset 20260827-2142-2f1b575
Download the run: manifest · per-suite counts · every verdict, columnar JSON
Both rates are here because either one alone hides something. The strict rate counts only tests the engine can be judged on. A test that needs JavaScript or a mouse cannot be passed or failed by a PDF renderer. The raw rate counts the whole corpus, every skip against us. The honest figure is somewhere in between. How the run works.
The same run, read four ways. Pick whichever matches what you came to check.
Two of them measure different things, and it is worth saying which. The test suite is self-contained: a test file and its reference are both rendered by our engine and compared, with no browser anywhere in the loop. It answers whether the engine agrees with the specification’s own definition of correct. The property atlas is the opposite: every example is drawn twice, once by us and once by Chrome, so it answers whether we agree with an outside implementation. Neither question subsumes the other, which is why both are here.
23,994 reftests across 81 suites. Not all of them can be pointed at a PDF engine. The ones that cannot are listed by name, and that is the difference between the two rates above.
Measurement ties differ only by anti-aliasing at the raster resolution. They count as failures.
Every excluded test, by nameThe six specifications the engine loses most often, among those with at least sixty judged tests. Each links to its failing tests and their diffs.
The web suite has almost nothing on paged media, because browsers rarely print. Running headers and footers, footnotes, leaders, named strings, cross-references. We wrote 98 reftests of our own for it, in the same format and judged the same way. They are not counted in the figures above. We wrote them ourselves, so they are a regression net; nobody should score an engine on tests its author picked.
98 of 98 passing · excluded from the headline figures
The file itself has standards to meet as well. Those are checked separately with veraPDF, the reference validator, over every built-in template.
A validator tells you a file does not break the rules it can express, which is a smaller claim than the file being accessible. Our own structural audit exists for the gap: it is what caught a PDF/UA-2 file we shipped with a %PDF-1.7 header that veraPDF passed, because that profile checks prohibitions and not the container version. The same audit found eleven structure-namespace violations and a footnote type that changes name between the two editions. All of it is written up, with the fixes.
We could not find per-specification test results published by any other HTML-to-PDF engine. Prince, PDFreactor and Antenna House publish feature lists; the most recent public conformance marker for this class of tool is Acid2, from 2005. That is the reason this section exists in the form it does, and if one of them publishes a comparable run we will link to it from here.
It carries our render, the reference, the diff between them and both source files. Suite pages list their tests with the size of that difference.