Technical-document evaluation pack - October 17, 2026

Nine document-reasoning tasks use selected text and visual excerpts from two
NASA Systems Engineering Handbook editions. Test table and diagram reading,
revision references, conflicting evidence and questions the supplied inputs
cannot answer. This is a small systems-engineering document corpus, not an
equipment-maintenance or operational military validation.

Start with DATASET-CARD.md and RIGHTS.md. NASA is credited as the source;
NASA does not endorse Qwerty. Full handbooks and excluded third-party material
are not redistributed. Source URLs, edition/page locators and hashes are in
sources.json. The reference review is disclosed AI-assisted editorial work.

Use Python 3 with no additional packages:
  python3 corpus-tool.py validate
  python3 corpus-tool.py report --help

Provide only sources/ and the relevant questions/Mxx.txt to the system under
evaluation. Keep tasks.json and REFERENCE-REVIEW.json away from its context.
Keep actual outputs and reviewer records outside this reference directory.
The checker validates citations and record structure; it does not determine
semantic correctness. Actual model outputs and acceptance measurements are
not included. Follow the dataset card to record a real evaluation.

SHA256SUMS.json covers the 26 source files in this edition.
Report source/key/rights issues through https://qwert.ai/contact with the
corpus ID, task, edition/page, file hash and proposed correction.