261 commits to Bow & Bridge from Mon 17 Aug to Sun 23 Aug 2026. The OCR pipeline read 12,000 pages this week, and most of what it got wrong was a name that wasn’t one: an author listed beside a book about makers, a dash where a forename should be, a middle initial read as a death date. I spent the week writing those refusals into the miner as rules, each with a guard that proves it still holds, and building the review screens where a person decides what is left.
Pipeline figures for the week: 12,096 pages read, 769 volumes unlocked, 19 records published and 3,043 corrections added.
- Wrote a set of name rules into the miner: an author is not a maker, a dash is not a forename, a middle initial is not a death date, a courtesy title or a printed place never sits in the name slot, and one printed line naming two makers becomes two mentions.
- Verified 2,893 claims that quote a printed line directly against the page text, with no language model involved.
- Retired 126 junk drafts (index lines, death notices, catalogue prices) that had been minted as makers, and deduplicated 1,307 byte-identical mention rows across 40 records.
- Built the review surfaces: one-click verdicts under the transcript viewer, a Flags list where a reader can dispute a misread page with a reason, and a same-man/different-man question answered from one popup.
- Added a dealer record type with evidenced person-to-person relationships, and turned the rights field into a corpus-wide dropdown that states its basis.
- Closed 61 tracked issues in one overnight run and 51 more after review at the weekend, and fixed three guards that had silently not been running for days.