dsh-skill-pdf-math

A DeepSeek Harness skill for reading mathematics out of PDFs and formula screenshots without guessing.

Why

Vision models misread formulas in predictable ways: a \mathcal{L} becomes \angle, a summation subscript j=1 becomes \cdot, a tall bracket turns into [[[. Most of that is unnecessary, because a LaTeX-produced PDF stores formulas as positioned text, not images — glyphs, baselines, sizes and font names are all there.

So the skill teaches one rule:

read the text layer first, let formula OCR supply the structure, and cross-check the two.

What it adds

• pdf_formula.py — turns a PDF region into readable (approximate) LaTeX by restoring sub/superscripts from baseline offsets, plus a small TeX cmex10 operator table.
• --probe — decides whether a page's formulas are extractable text or images.
• --crosscheck — compares the OCR result against the text layer's letters and their fonts (cmsy10's L is \mathcal{L}), so a wrong glyph is named instead of merely suspected.
• --verify — re-recognises with a different crop and resolution; disagreement is reported.

Measured on four numbered equations from a real paper:

render	mean similarity	exact
400 dpi	0.941	2/4
600 dpi	0.977	3/4

--check requires the optional rapid_latex_ocr runtime; without it the text-layer channel still works and the skill says so instead of failing.

Install

dsh plugin --profile <name> add dsh-skill-pdf-math

or install it from the DSH plugin marketplace. The skill registers itself as pdf-math and is model-invocable.

Notes

• Registration is wrapped in a try/catch: a broken asset can disable the skill, never the host application.
• Paths are injected at run time, so the published package contains no machine-specific paths.

License

MIT