D30: The rendered subset of LaTeX

200 papers per field submitted in 2024, five fields, an even share of five categories each, taken in a fixed order so that every run measures the same papers: mathematics (math.AG, math.PR, math.CO, math.AP, math.NT), physics (hep-th, cond-mat.str-el, astro-ph.GA, quant-ph, physics.optics), computer science (cs.LG, cs.CL, cs.DS, cs.CR, cs.CV), biology (five q-bio categories) and economics (econ.EM, econ.TH, econ.GN). Papers without a LaTeX source are skipped and the next ones taken. arXiv's sources stay out of the repository: most are not redistributable. The development containers cannot reach arxiv.org, so the study runs in CI; the sample it fetched is kept as the assets of a draft release, which only the repository's writers see, so that a run does not ask arXiv for it again (the Actions cache evicted it) and a writer can measure it locally.

The body of every .tex file (\begin{document} to \end{document}, or the whole of an included file), without the files a document brings in before \begin{document} (its macros): the bytes the view shows as source, the bytes in formulas (the math renderer's), and every command and environment, ranked by the papers using it and by the bytes it shows as source. "Shown as source" is what the view shows: the outermost command it does not render, whole; for an environment it does not know, its \begin and \end (the view shows the body as text, each thing in it as it shows that). The arguments the view draws or hides (a picture's options, a label's key) are not text shown as source.

The first run counted otherwise: the body of every unknown environment whole, the arguments of drawn commands, macro files, and a list of rendered commands that had fallen behind the view. It put 19.2% of the body as source, with \frac and \textwidth near the top. The metric was corrected to what the view shows before any number was used.

Since run 9 a formula counts as source when the math renderer cannot read it, prepared as the editor prepares it (the document's definitions, the packages' commands the renderer lacks, references as their numbers): the view shows such a formula as its source in a red frame. Since run 25, when TeX is installed, the view has it typeset such a formula on its own (a standalone document with the paper's preamble) and draws the result in the text's color, as it draws TikZ pictures. The report counts these formulas apart ("by TeX") with the renderer messages behind them: they are what a reader without TeX still sees as source, and what the renderer's preparation should learn next. The files no document of the paper reads (old versions, unused sections) are left out.

Run 36 of the workflow (2026-10-03, commit 2603aa5): 925 papers, 2,496 files, 86 MB of body text.

Field	Papers	Files	Body KB	Source	Formulas	By TeX
biology	189	588	14,375	0.05%	10.65%	0.03%
computer science	179	869	16,220	0.09%	12.69%	0.34%
economics	183	484	22,225	0.06%	22.87%	0.15%
mathematics	187	256	18,713	0.06%	39.30%	0.45%
physics	187	299	14,646	0.09%	16.97%	0.52%
all	925	2,496	86,181	0.07%	21.48%	0.29%

Every field is under the target. Run 30 (2026-10-02) had 0.15% (0.08% to 0.26% by field), run 17 had 1.67% (0.64% to 2.54%). What shows as source now is the long tail: each paper's own macros that neither print text nor draw what a reader can follow, packages' environments met in one paper, tables whose columns are formulas (>{$}l<{$}), quantum circuits (\Qcircuit).

The runs between 30 and 36 added: files not in UTF-8 (inputenc with latin1 or applemac; one such file lost a paper all its theorems), equation macros a project's other files define (\def\ba{\begin{eqnarray}} in its package), a document's own macros that are references, space, displayed formulas or boxes, or that draw pictures and read files (twoopt's included), \ifmmode's text branch, the usual short names of theorems, algorithm2e's statements, aastex's tables, words and front matter, REVTeX's, ICML's and amsart's front matter, A&A's \ion, TeX's conditionals in the text, and picture, diagram, xy, ytableau and overpic drawn by TeX.

The runs between 17 and 30 added, from their traces: the root's definitions in the files it includes, macros that print their argument, macros with arguments that make text, macros that define macros, macros that are formulas, equation macros, letter families, glossaries, pifont's and wasysym's symbols, TeX's parameter settings, the lists of enumitem and paralist, algorithm2e's settings, and formulas TeX typesets when the renderer cannot read them.

Run 9, the first to count formulas, had 6.4%. What shows most as source now: a document's own macros a formula finds undefined (a long tail, one or two papers each), commutative diagrams (tikz-cd, xy), young tableaux, TikZ pictures, plain TeX in formulas (\aligned, \buildrel, \matrix), formulas between macros for an equation (\be … \ee).

Run 7, before formulas were counted, had 0.6% (0.4% to 0.9% by field).

Rendered, each checked against pdflatex by the differential tests of T2.7h.33 (tools/latex-typeset-fuzz.py, tools/latex-numbering-fuzz.py, tools/latex-edit-fuzz.py, in CI):

• Structure: sectioning with numbers as the class numbers them, \appendix, front matter and \maketitle, the abstract, \tableofcontents.
• Text: the font commands and declarations, sizes, accents and the text symbols of LaTeX and textcomp, quotes, dashes and ligatures as TeX makes them, spacing commands, case changes, \verb, colors (\textcolor, \color), boxes (\mbox, \resizebox, \parbox, …).
• References: \ref and its kin with the numbers LaTeX gives (counters, \numberwithin, amsmath's equation labels and their clashes, enumitem's labels, subfigures, \captionof), citations in the style of \bibliographystyle or natbib, and a document's own thebibliography (REVTeX's included).
• Blocks: lists, floats and their captions, theorems as \newtheorem declares them (in the document or a package beside it), tables, displayed formulas.
• Markup that prints nothing where it is (\newcommand, \def, \setlength, \captionsetup, …): dimmed.

Drawn by the TeX installed, when there is one: TikZ and pgfplots pictures, and formulas the math renderer cannot read; without TeX they stay as source (pictures folded). Left as source, for now: unknown commands of a document's own whose definitions are not read, and environments the view does not know (their delimiters).

The first target, 97%, was set when every field was above 99% by the view's commands alone. Counting formulas showed what a reader meets: a formula in a red frame is as unread as a command shown as source, and mathematics, where a third of the body is formulas, was the field furthest from the PDF. The target is a thousandth of the body in every field, so that a paper opened in Kalem reads as its PDF does; a fall below it is a regression in the view, the renderer's preparation or the metric, found by the monthly run.