Bow & Bridge
Bow & Bridge810
- 20267 weeks logged810
- W3921 Sep15commits
Bow & Bridge
A facsimile is not a second edition
- Added /labels/: every checked maker's label on a published record, A to Z, with the book, scan page and line it was read from.
- Fixed the edition rule: a facsimile counts as the edition it reproduces. 105 records resting on one reprinted book were demoted to draft, and 583 same-person questions were answered in the same pass, leaving 569 published makers.
- Fixed the OCR loop check, which had rejected 1,019 honest pages across 429 volumes because a row of dot leaders looked like a decoding loop.
- W3814 Sep74commits
Bow & Bridge
First production deploy: 605 makers live
- First production deployment on 18 Sep, second and third on 20 Sep; the deploy is one command, dry by default, ships the tables as data and rewrites URLs on the host.
- Applied a ninety-year rule: a maker whose cited printing is at least ninety years old clears the possibly-living check. The export went from 191 to 648 published records in one pass; 605 were live after the second deploy.
- Site review for a trade reader, eight findings fixed: search finds makers, every published maker in one A–Z table, the citation stays open on every page, each printing links to its scan.
- W377 Sep34commits
Bow & Bridge
Fifteen ceilings the pipeline must stay under
- Fixed the admit storm: one volume the station could never fit had been asked 10,506 times; the admission rung is now paced by work done, and the drain front rotates instead of starving.
- Wrote the operations contract: fifteen ceilings (queue depths, run times, log sizes) read from one file, with a nightly uncapped guard pass and a station ledger that records every page ask.
- Stopped the miner minting bare surnames: a single-token name is refused, and a guard catches any that got through.
- W3631 Aug155commits
Bow & Bridge
An admission door for every book, and a site a stranger can read
- Built the admission door: a volume is sampled, profiled and given a rule-table verdict before any mining; the miner, the prose stage and the import step all refuse a volume the door has not admitted.
- Rewrote every public page and template string for a general reader, rebuilt the maker index around search with a random cut of 18 records, and added a public /progress/ page that counts the pipeline stage by stage from a snapshot.
- Closed every code path that could mint a second live record for the same name, then restored 177 merges that an August rebuild had silently undone.
- W3524 Aug85commits
Bow & Bridge
Two verification bugs that hid 3,500 good claims
- Fixed the verbatim verifier's apparatus detector: attested claims went from 0 to 3,545 in one run.
- Fixed page-locator resolution in the batch verifier from 0.3% to 99.0% of claims landing on their exact page, and stopped it spending model calls on retired records.
- Indexed the transcript table so a per-claim lookup dropped from a 92-second full scan to 2 seconds.
- W3417 Aug261commits
Bow & Bridge
Teaching the miner what is not a name
- 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.
- W3310 Aug186commits
Bow & Bridge
Bow & Bridge goes live as a sourced reference site
- Stood up the site: maker records where every claim carries its own citation, an instrument register, source pages, a citation block on each record, and llms.txt and robots.txt so crawlers and AI tools read it correctly.
- Read the complete 1916 American directory of makers from page images (412 entries) and published it as evidence rather than biography, each entry linked to its printed page.
- Ran an adversarial review of the published records (3 critical, 12 major, 21 minor findings) and re-read 79 published claims against their sources; four unsupported claims were corrected and five withdrawn, with the corrections filed on the site.