# Historical body-text development review

This step prepares annotation work, not another model-quality score. It uses
paragraph regions (including verse) from the pinned validation split of
PiotrSty/impact-psnc-polish-ocr, excluding the three previous heading-control pages.
Choose up to four regions per collection, round-robin across pages, in stable ID
order. Selection does not depend on predictions or detector success.

```bash
python -m training.prepare_body_dev_review --output data/body-dev-review-v1
```

Requires Pillow, numpy and OpenCV; no GPU or model weights. Original JPG crops,
source metadata and SHA-256 hashes are retained. PNG conversion preserves RGB
pixels; line crops use the original RGB image without deskew or binarization.
OpenCV normalization is used only to propose geometry, not alter saved pixels.
Proposals preserve the full region width and divide vertical gaps at their
midpoints; the first and last crop extend to the region edges. This avoids tight
ink-band clipping but can retain neighbouring fragments, requiring visual review.

## Review

1. Open `region-review/index.html` to inspect original paragraphs and references.
2. Open `line-review/index.html` to inspect proposed individual line crops.
3. Check the entire line is present, no adjacent text is included, and the
   reference really belongs to that crop. Preserve historical spelling.
4. Mark doubtful geometry or glyphs as requiring review. Do not silently replace
   private-use symbols or illegible text with a model prediction.
5. Export the JSON review history. Use independent reviewers/profiles before
   treating agreed labels as a candidate reference set.

The existing panel calls records pages; here each record is a region or line.
All labels begin unverified. The line/reference alignment is a count-only
proposal, not a certified correspondence. Models are not run or shown to reviewers.
Count mismatches and overlapping boxes are logged in `report.json`, with their
source regions still available for review; no reference strings are assigned to
these line proposals. Do not discard these failures from a future full-page score.

## Boundaries

`candidate-lines.jsonl` is NOT an evaluation or training release. Every record
has `eligible_for_evaluation: false`; the generic review tool does not turn that
flag on. Review/adjudication and a separate audited release step are required.
Do not feed this file straight into an evaluator that ignores those fields.

Frozen test collection IDs and exact crop hashes are excluded. Near-duplicates,
shared editions and upstream training overlap are not ruled out. In particular,
similar collection names must not be assumed to identify independent works.
These are historical paragraphs/verse, not a representative sample of modern
Polish documents. Check typeface coverage manually before expanding or training.

Raw images, draft references and review histories remain local. Only preparation
code, tests and methodology belong in the repository until publication is approved.