Registrum B (Hernando Colón's catalogue of printed books): machine reading of the 1905 facsimile, version 3.8

Machine reading, human check pending. This record holds an AI-assisted reading of Registrum B, the numbered catalogue of printed books of Hernando Colón's library (nos. 1 to 4231), made from the 1905 Huntington facsimile (digitised by Google, Internet Archive; public domain in the US). No person has checked the readings against the scan. Agreement between machine sources is not accuracy. This is not an edition. Contents: a table of 4,205 entries (CSV, 36 columns), the same entries as TEI P5, a SQLite database, an offline search page, lists of changes against versions 3.5 and 3.7, a list of unresolved entries, a changelog, a README and a build report. 26 catalogue numbers are not in the table. What version 3.8 changes against version 3.7: Two more page-columns (92L and 181R) had their box numbers attached one entry too high, as in the 61 columns corrected in v3.7. They are corrected: 63 corrected columns in all, 523 renumbered entries, 63 merged continuation rows, 54 entries added from the scan. The two new entries (1325 and 2639) carry only the first words read from the scan, so 70 entries are "first words only" (v3.7: 68). The table still has 4,205 rows. 69 entries on pages once believed unreadable (92R, 181L, 274, 279R, 286) had been flagged only for that reason. A machine re-read of the Internet Archive PDF found their box numbers and first words in agreement with the table, and the flag is withdrawn. Entries flagged: 18 (v3.7: 107). The agreement checks were recomputed: the Abecedarium B supports the entry number in 3,447 of 4,205 entries (82.0%; v3.7: 3,431, 81.6%); against the Madrid copy (BNE MSS/18119), in the two samples of entries, 231 are aligned, 20 shifted and 19 without match (n = 270; v3.7: 230, 21, 19). These are agreements between machine readings, not accuracy. New in the CSV, SQLite and TEI: genre (machine rule layer v3), book language (model label), a language heuristic, and normalised place, place-is-purchase flag, year, decade and format. All are machine labels carried over from the earlier table by entry; they are empty for the 70 first-words entries (genre "UNCLASSIFIED_NEW"). For entries merged with a continuation row they come from the first part only. The two language fields disagree on many entries; the book language is the one used in our articles. New column v3_7_registrum with the number each row had in v3.7. Limits: every field is a machine reading and has not been checked by a person. The corrections of the box numbers are machine re-reads of the scan (two columns of v3.8 by one reader). 209 entries have no title; 26 numbers are missing; 18 entries are still flagged. Deletions, insertions and corrections of the manuscript are not recorded. No ISTC or USTC numbers are assigned and printer names in edition suggestions are unverified. The TEI file is well-formed and passes our integrity checks. On 30 October 2026, after this version was built, it was validated against the official tei_all.rng (TEI P5 4.12.0) with lxml RELAX NG validation: valid, 0 errors, on 4,205 bibl entries. This shows conformance to the schema, not that the encoding choices are sound; no TEI specialist has reviewed them. README_v3_8.md and CHANGELOG_v3_7_to_v3_8.md in this record were written before this check and still say that the file has not been validated. No data from the Catálogo concordado, the Colombina catalogue or CERL is included. Correction note (30 October 2026). The README in this version (README_v3_8.md), in the section "Known limits", says that of the 61 columns corrected in v3.7, 24 were read by two different models. This is wrong. The correct figure is 12 columns read by two different models (a second pass also read 12 unshifted control columns); the other corrected columns were read by one pass. A corrected README will be included in the next version. Published files cannot be changed in this record.