# Documentation clarification — 2026-11-11

- Distinguish 17,285 released training records, 33 export exclusions, 17,252 retained records and the 17,995-row mixture including 743 auxiliary records.
- Distinguish rank 128 per training run from the exact averaged rank-256 inference representation.
- Clarify that existing cost estimates use earlier rank-16 timing; no new timing, training, weights or scores.
- Describe the OwlEye correlation as an association, not proof of localization transfer.
- Original dataset Parquet files and the immutable eval-2026-11-10 tag are unchanged.

# eval-2026-11-10

Additive evaluation revision for arXiv:2611.10234.

- Preserve original train, validation and test splits, labels and published main leaderboard.
- Add base-page and requirement-group bootstrap intervals from frozen predictions.
- Add unfiltered, author-audit-excluded, structural-quarantined and convention-sensitive views.
- Add a same-model inference-cut ablation and output-format/family diagnostics.
- Disclose model-dependent exclusions and annotation limitations; no full relabelling, new training or independent human validation.
- The arXiv link currently points to the announced paper; this revision may precede its next arXiv version.