# Deviations from the pre-registration

1. **A fifth class, H (handwriting).** The pre-registration defined a dropped line as printed text
   and gave four classes. Three sampled items are manuscripts (a Karnataka state archive file,
   `karnataka-state-archives-RGVlcGFrMTM5MzQz-RGVlcGFrMQ`; a Baptist meeting's minute book,
   `per_early-baptist_whites-alley-meeting-ho_march-03-1699-july-26-1708`; and
   `BANM_ARC44d1_015`, whose only flagged ink is a handwritten margin mark). The classifiers,
   told only the four classes, called most of these bands A. The author moved all 17 to H, which
   counts toward none of the outcomes. Counting them as A would add two pages (the Karnataka file
   and the minute book) to the 26.
2. **Which bands the author re-examined.** As registered: every A or B call, and a seeded 10% of the
   C and D calls (36 of the classifiers' 359, seed 20260929). All 86 final A calls were re-examined.
3. **A second witness not in the pre-registration.** After classification, tesseract (English
   model) read every re-examined band off the image, and each word of five letters or more was
   looked up in the page's whole text layer (`adjudication/tesseract-witness.jsonl`). It decides
   nothing and changed no call; it is reported as corroboration.
4. **Two printed pages with an empty text layer** (`india.history.resource.6852` and
   `bim_early-english-books-1641-1700_the-parallel-or-a-compa_1682`) are counted as pages with
   dropped body text and left out of the per-line rate (outcome 3), since they have no line count.
   The pre-registration did not anticipate an empty layer on a page with text. (A first draft of
   the page treated only the first this way; the claims pass caught it. Two further usable pages
   with empty layers are blank images.)
8. **G3's first half** ("most C candidates are rules, borders and show-through") was not scored
   band by band. The page reports a keyword count over the classifiers' notes instead.
5. **The search-inside probes** (`search_inside.py`) were added after the census, for the page, and
   are not an outcome.
6. **"Kind" (newspaper or periodical vs other)** is a keyword match on collection, identifier and
   title (`analyze.py`, `kind`), widened once after reading the list of usable pages, because the
   first pattern put titled newspapers (L'Opinione, The Bombay Chronicle, El Día) under "other". It
   is still approximate (Adventure magazine and Głos Prawdy fall under "other").
7. **Pilot claim checked late.** The pre-registration says a pilot page was page 317's image with
   page 318's text. That was inferred from the hOCR when written; the image was looked at on
   2026-09-29 afterwards and it is headed 317.