論文の要約（数値はページ・節つきで引用）

What it does

Summarises a paper so that every number and claim points to the page and section it came from.

• Reads a PDF (via pdftotext) or extracted text and detects page breaks and headings (numbered headings and names such as Abstract, Introduction, Methods, Results)
• Labels every sentence with a reference like (p.5, §4.2 Results)
• Tabulates sentences that contain numbers (%, plus-minus, p-values, counts, K/M units), skipping table-like garbage and the reference list by default
• Picks candidate sentences for problem, contribution, method, data, results and limitations by keyword, preferring the matching sections
• Lists table and figure captions with pages
• Ships a summary template: problem, method, data, results, limitations, follow-ups

How to use

python scripts/paper_sections.py paper.pdf --outline
python scripts/paper_sections.py paper.pdf --out grounding
python scripts/paper_sections.py paper.pdf --section 4.2

The agent reads grounding.md and the key sections, then fills the template, quoting every number with its page and section.

Requirements

• Python 3 (standard library only)
• pdftotext (poppler) to read PDFs directly; otherwise pass extracted text (the script stops with a clear message if it is missing)

Caveats

• Heading and sentence detection is heuristic and can slip on two-column layouts; check with --outline.
• Table cells are often scrambled by text extraction, so the procedure takes table numbers from the original PDF and cites the table and page.
• Scanned PDFs need OCR first.