Research question
Now that MathML Core ships in every major browser engine, what actually happens to the equations in public STEM and education PDFs when they are converted to HTML?
Overview
For the first time since the format existed, the display problem for accessible math on the web is solved: MathML Core renders natively in Chrome, Firefox, and Safari, and modern screen readers speak it. The remaining problem is upstream, in conversion: equations in PDFs are routinely dropped, flattened into garbled character runs, or rasterized into images with no alternative.
This study measures equation survival through common conversion paths on public STEM and education documents: dropped entirely, present as broken text, preserved as an image with or without alternatives, or converted to real MathML. It pairs the field measurement with an evaluation of the practical LaTeX-to-MathML toolchain for converters that recognize formulas.
Our position in this study is declared: our pipeline currently preserves tagged formula content as flagged text for review rather than dropping it, and full formula recognition to MathML is on our published roadmap. The study measures the whole field, including us, before and after.
Why this matters
STEM education runs on exactly these documents
Course notes, problem sets, and public research PDFs are the daily reading of students who use assistive technology. Every dropped equation is course content that simply does not exist for them.
The excuse expired
For years, poor browser support justified shipping math as images. With MathML Core native everywhere, the remaining failures are conversion failures, and measuring them is how they get fixed.
What we measure
- Equation survival
- Per-equation outcome: dropped, garbled text, image without alternative, image with alternative, or structured MathML
- Inline versus display
- Survival rates split by inline and display math, which fail differently
- Toolchain quality
- LaTeX-to-MathML conversion accuracy for the recognizer-plus-converter path
- Screen reader outcome
- Whether surviving math is actually announced correctly by current screen readers, on a sample
Method
Corpus: public STEM and education PDFs with equations enumerated and recorded as ground truth before conversion.
Subjects: common conversion paths and document AI parsers, versions recorded, plus our pipeline with its formula handling at its current stage, stated plainly.
Assistive technology sample: manual verification with current screen readers on a stratified sample of surviving equations.
Evidence so far
Only recorded findings appear here. Anything not listed has not been measured yet.
The failure mode is registered in our own stack
Until August 2026 our tagged-PDF path silently dropped Formula element content; it now preserves the text reading and flags it for review, with the defect and fix documented in the repository. That trajectory, drop to preserve to structure, is the one the whole field needs to walk.
Questions we expect
Why not just keep equations as images?
An image freezes math at one size and one rendering, and its alternative text can only be read linearly. Real MathML supports navigation within the expression, braille output, and reflow. Images are the fallback, not the goal.