v4.41.0

check-draft: the second check for a report that is already written. Where
check-references asks whether each reference is the work it names, this asks
whether each figure and quotation is on a page of the files the draft cites.

Added — check-draft: is each figure and quotation on a page of the sources?

report-workflow check-draft draft.docx --source a.pdf --source b.pdf checks a
finished draft (.docx, .md, .txt) against the files it cites, with no
workflow run, no network and no model (draft_check.py). Only sentences that
state a figure or quote are read. A figure is found when a page prints it at
the precision the draft gives — 69% is 68.9 rounded, 68.0% is not 68.9 —
with a measured unit that agrees (21 hours is not a bare 21.7) and within 250
characters of at least two of the sentence's own words, leaving out words on
more than half the source's pages (one word of any kind for a figure of four or
more significant digits printed exactly). A quotation is found when a page holds it
word for word, ellipses allowed, including across a page break. Each answer
names the file, page and the source's sentence; a figure not found names the
nearest one the page does print there. Citations, model names (LLaMA-65B, β1,
pass@1) and pointers (Table 3, 第 2 章) are not taken for figures. A PDF page is
read three ways — in stored order, one column at a time where the words leave
a gap down the page, and as parse_pdf reads it — so a two-column paper's
sentences are found whole. --appendix out.md writes the comparison as a page
to hand in, in Chinese for a Chinese draft, with each PDF page cropped and the
match highlighted under out_files/; --json writes every verdict; exit 2
when anything is not found.

numbers_with_spans in factuality_check.py returns FE's figures with their
offsets, and _page_text passes pdfplumber options through; FE and the
pipeline read exactly as before.

Measured — check-draft on three CC BY 4.0 arXiv papers

A correct and a planted draft for each of LLaMA (development), QLoRA and
Chain-of-Thought (a Chinese draft about the English paper); the last two were
committed before the command first ran on them. Held out: 8 of 141 correct
items flagged, 32 of 39 planted errors caught — every altered and invented
quotation, wrong unit and added decimal, 14 of 17 wrong digits, and none of the
four figures moved to another model in the same table. Development: 0 of 59
and 17 of 20. scripts/run_draft_benchmark.py --check reruns it from the PDFs
in benchmarks/check_draft/ against
benchmarks/evidence/check_draft_2026-10-21/.