Test case: Burgers equation identification PDF (256KB, mostly figures).
Result: 7/10 — correct qualitative summary, missed quantitative details.
Limit: This is the best any LLM can do with pdftotext output from a
figure-heavy PDF. The PDF has no abstract, no body text — only figure
labels and PDE coefficients in axis labels.
Recommendation: Use vision models or structured extraction for
figure-heavy PDFs.
