[yodablocks]

What

The 15-question hard set from #30 was labelled synthetic-draft (drafted by an LLM, not reviewed). Marc has now checked every answer and page of h-003..h-011 against the PDFs, so those rows are marked synthetic-draft+human-accept.

• Data: experiments/recall/hard_v1.jsonl. Provenance changed for h-003..h-015 (h-001 and h-002 were already accepted from the first gold set). Each row's note records the check. The file now loads with the default acceptance check (load_gold(path), 15 questions), so quarq eval --dataset accepts it. I did not run the eval: it loads two models on the GPU.
• Docs: a dated update in the recall README, and the Experiments caveat in the top-level README reworded.

What was checked

Each answer and page was opened in Preview and compared with the passage, using the PDF viewer's title-bar page index. In some of these files that differs by one from the printed page number (h-008: viewer page 7 is printed page 6), and quarq eval requires the viewer index.

The four French rewrites (h-012..h-015) share their English originals' answers and pages and were checked through them. Their French wording was not separately reviewed.

Caveats

• The LLM still chose the questions, so drafting bias remains. Only the answers and pages were verified by a person.
• The earlier "not human-reviewed" caveats in the feat(experiments): Jev as the re-ranker, compared with bge on the gold set #28 and feat(experiments): bring the recall experiment (#29) onto master #30 PR text and in the results sections of the recall README describe the state at the time and are left as written.