W9 address pour — extract-v1 over the W9 corpus: 233 figure / 67 event / 550 place / 92 people candidates offered (existing rows untouched), 18,240 attestations
receipt:addresses-w9 · v1 · ingest_verified
sha256 c3f254e44311166a10d5e6d503bb837e5751ac32746b646e210a94611c3b1a83
A receipt records what the house consulted and what it did not establish. The open state is part of the record, not a defect in it.
- Apparatus shelf + pinned Wikisource — Livy, The History of Rome (Rev. Canon Roberts translation, Everyman's Library)Rev. Canon Roberts, Everyman's Library (J. M. Dent & Sons / E. P. Dutton), first issue 1912; six volumeslicense: public-domain (the Roberts translation's Everyman first issue is 1912, pre-1930; Wikisource dates the translation 1905 — either way decades inside the US public domain; digital-door text carries no additional rights)pinned: TWO DOORS, one translation, stated honestly: (1) shelf djvu scans vols I-V — 'Livy - The History of Rome [Vol I..V].djvu', sha256 2e02ebff590ebaf6…/3c4df8aaeeffb255…/3eb73eb49c30dba2…/2e5b7572db51e439…/b21ec32a66a5fb99… (full hashes in ops/sources/books.inventory.json), text layers re-ordered from their own geometry by profile livy-roberts-v1 (ops/pipeline/profile_livy_roberts.py); (2) pinned en.wikisource.org digital transcriptions of the same translation — Book XXXIX (revid 3795019, sha256 f1a3ada49b10c956…) and Books XL-XLV (revids 3795016/3795012/3784681/3795009/3784683/8966066, sha256 012d7b73fa2f78c9…), acquisition pins in ops/sources/MANIFEST.md; plus 55 whole-chapter cures for 3 missing print leaves and the OCR holes in the scans (per-chapter list frozen in livy.works.json and livy.leaf-cures.json; witness revids in livy.witness.json). Books 11-20 are lost in the manuscript tradition (only the Periochae survive). Chapter boundaries were verified against a committed boundary witness built from the same pinned door; canon chapter counts double-sourced (critical editions + the witness, 35/35 books agree)
Deterministic extract-v1 (unchanged prototype logic, W9 corpus scope) over the W9 episodes. 18,240 attestation rows carry exact surface forms and per-episode counts (receipt_id = receipt:addresses-w9 on the rows themselves). Existing addresses gain W9 attestations by slug identity and have their mention_count recomputed from the full attestation ledger; their curation-owned columns (labels, kinds, notes, states, indexability) are untouched by construction — the insert is ON CONFLICT DO NOTHING and the only UPDATE touches mention_count. New candidates are born candidate/unreviewed/noindex. The W9 unclassified pool and era phrases stay local (D-LEXICON-001; era is book-gated W3 law).
What this does not prove- extract-v1 is frequency + pattern extraction over partly OCR-damaged text — not NER, not curation; OCR errors propagate into candidate labels and surface forms, and misfiled rows exist at the margins, this batch's named: 'Antiochus' (376 mentions) is filed as a PLACE candidate by the locative-context heuristic (war-narrative contexts), 'Roman' as a place, and 'Dictator', 'Carthaginian', 'Gauls' and 'Samnites' as figure candidates — the same class of margin W1's 'Roman' sat in, awaiting curation lanes.
- This batch's extraction scope is the W9 work alone (extract-v1's gates are corpus-relative; W1/W6/W6b/W7/W8 were poured first and stay append-only) — candidate counts are not comparable across batches.
- Attestations attach to existing addresses by surface-slug identity, not by curation: 'Hannibal' in Livy attaches to the same rows the Polybius and Nepos-era waves fed, and Livy's many homonymous Romans (the Scipios, Fabii, Claudii) attach to shared surfaces — homonym mixing is possible on any shared surface.
- mention_count is corpus-coverage arithmetic (full-ledger sum for figures/places/peoples the wave touches); event mention_counts of pre-existing events keep their original wave's scope.
- figure_kind is 'unassigned' on every new candidate — person|legendary|deity is a curation verdict.