NuBerea/nt-apocrypha

New Testament apocrypha — critical editions, region-grain OCR

Three public-domain critical editions of the NT apocrypha as region-grain OCR, with the index / title / running-head / marginalia / apparatus registers separated from the body text. corpus='apoc'.

volume_id	edition	year
APOC_001	Tischendorf, Evangelia Apocrypha (2nd ed.)	1876
APOC_002	Tischendorf, Apocalypses Apocryphae	1866
APOC_003	Lipsius–Bonnet, Acta Apostolorum Apocrypha	1891–1903

Configs

config	what
pdf	pointers to the source scans (repo + path + sha256) — the scans themselves stay in the source archive and are not copied here
ocr	per-region OCR — volume_id/page/region_idx/anchor/region_label/lang/bbox/text/conf/provenance/match_grain/osisRef
ocr_challenger	the same regions with per-engine text + confidence (Tesseract vs Kraken), so engine disagreement is auditable rather than silently resolved
works	derived per-work main-text views (body registers only), keyed back to ocr; 56 rows (20 APOC_001 / 11 APOC_002 / 25 APOC_003, work_spans_v2/); text is the raw glyph-faithful engine reading and text_normalized applies the edition glyph map — both are shipped, neither is lost
repair_sidecar	the 41 reviewer-accepted additive text repairs bound to the v4 anchors (32 carried candidates + 8 re-anchored from the prior stage + 1 two-reviewer-accepted region re-OCR); each row carries source_unit_id, before/after text + sha256, method, status, and review receipt — raw ocr is untouched
structural_role_sidecar	4,061 region-role corrections: 78 v9-shipped rows (77 page-number stamps excluded from the derived claims feeder + 1 ink-adjudicated relabel restoring a main-text paragraph the zoning had filed under the apparatus) plus 3,982 index/TOC-page regions across all three volumes relabelled Marginalia_IndexTOC and excluded from the claims feeder (36-row evidence sample, 0 wrong) — raw OCR preserved in every case
confusion_sidecar	the lexicon-gated, ink-FP-receipted κ/χ correction layer for APOC_003 (1,019 regions; raw untouched, substitution list + gate per row)
apparatus_entries	the critical apparatus decomposed into individual lemma/variant/witness-sigla entries, both editions combined (27,049 rows)

Per-volume text-bearing rows (v11, re-derived from the shipped ocr parquets)

volume_id	total ocr rows	− Marginalia_ScanBleed	− ScanBleed + Marginalia_LineNumber	body-label rows (MainText_*)
APOC_001	2,957	2,951	2,950	1,303
APOC_002	1,036	1,028	1,028	377
APOC_003	18,582	11,318	9,514	5,096
corpus	22,575	15,297	13,492	6,776

Body-label rows are the region_labels works draws from (MainText_ColGreek / MainText_ColLatin / MainText_Title); the gap between the ScanBleed+LineNumber column and body-label rows is apparatus + running-head + other marginalia, not lost text -- every row is still in ocr under its own label. Owner commitment 2026-08-28 (APOC convergence correction #4).

Region labels

MainText_ColGreek / MainText_ColLatin / MainText_Title — body text; Apparatus_R1..R3 — the critical apparatus registers; Running / Marginalia_PageNumber / Marginalia_ParagraphNumber / Marginalia — page furniture; PAGE_FAILED — a page whose OCR did not land (sentinel row, never silent loss).

APOC_001 works (20)

Tischendorf 1876 covers the whole volume as 20 works (boundary closure via the printed conspectus INDEX pp.XCIV-XCV plus 19 title-leaf renders; 332 previously work-less body pages closed to 0). Page numbers below are the printed range; printed_range in works carries the same.

work_id	lang	printed range
protevangelium_jacobi	grc	1-50
pseudo_matthaei_evangelium	lat	51-112
evangelium_de_nativitate_mariae	lat	113-121
historia_iosephi_fabri_lignarii	lat	122-139
evangelium_thomae_A	grc	140-157
evangelium_thomae_B	grc	158-163
evangelium_thomae_latinum	lat	164-180
evangelium_infantiae_arabicum	lat	181-209
acta_pilati_graece_A	grc	210-286
acta_pilati_graece_B	grc	287-332
gesta_pilati_latine	lat	333-388
descensus_christi_ad_inferos_latine_A	lat	389-416
descensus_christi_ad_inferos_latine_B	lat	417-432
epistola_pilati_altera	lat	433-434
anaphora_pilati_A	grc	435-442
anaphora_pilati_B	grc	443-449
paradosis_pilati	grc	449-455
mors_pilati	lat	456-458
narratio_iosephi_arimathiensis	grc	459-470
vindicta_salvatoris	lat	471-486

(11 lat / 9 grc. APOC_002 = 11 works, APOC_003 = 25 works; both use the printed conspectus + title leaves the same way and are not re-listed here — see works.parquet.)

Method

Split-first pipeline: each page is decoded from the source JP2 and OCR'd independently, with the page JSON as the sole checkpoint. Layout is deterministic zoning (apoc-zoning-v3), not a learned detector — the Migne-trained YOLO does not fire on these editions. The apparatus is cut off the body by a printed short-rule detector where the edition prints one (Tischendorf), and by an OCR text-signature fallback where it does not (Lipsius–Bonnet). Each Greek region is read by both engines; the primary differs by volume. APOC_001 uses the independently reviewed, Tischendorf-face Kraken finetune (model sha256 10a5499d2c96b09002c505f189c9f02d9edf6861d546a026958fc4f56c18dc95); APOC_002 and APOC_003 use the Migne-trained Kraken recognizer. The original Tesseract-led APOC_001 output is preserved unmodified and never overwritten.

Granularity (preserved)

The region-grain ocr table is the source of record; works is a derived view over it, never a replacement. Every text row carries match_grain and an osisRef anchor column so verse-level structure can be reconstructed later. The engine reading is never overwritten: where an edition glyph map is needed to make the text readable (the 1876 face prints ν as μ and has no final ς), it is applied into a separate text_normalized column beside the untouched text.

Known limitations

Each caveat below is updated in place as later verification supersedes it, with the supersession noted inline where it applies. Not claims-ready.

• OCR quality (dictionary-hit, same method and lexica used across the estate's other patristic OCR corpora). On a stratified sample (8,215 of 77,174 rows, seed 42; APOC_001 Greek body at full population, 948 rows) with the engine in the stratum key: Latin body 0.8247 (primary) / 0.8207 (Tesseract challenger) vs the Patrologia Latina corpus 0.876 / the Scholastics corpus approx. 0.884 (both post-repair); Greek body 0.7286 (primary) / 0.7572 (Tesseract-only challenger) vs the Patrologia Graeca corpus 0.7895 -- below the Patrologia Graeca baseline for every primary stratum, with one exception: APOC_002 Tesseract challenger Greek body 0.7919 edges above it; critical-apparatus regions 0.6982 / 0.6999, reported separately because sigla and abbreviations are not prose. These are raw/lightly-repaired figures (83 accepted repairs); no Greek clean-text control exists (the Latin word-shape method is not reusable for Greek). Dictionary-hit is lenient on short garbage tokens, so absolute rates overstate legibility while engine-to-engine gaps are fair.
• APOC_001 (Tischendorf 1876) Greek primary is a face-specific Kraken finetune. Against the Tesseract challenger on the same units it scores 0.6937 vs 0.6644 dictionary-hit on the sample (+2.93pp) and 0.6965 vs 0.6553 at full population (+4.12pp); a 24-case blind paired human review preferred the finetune 19:1. On APOC_003 the opposite holds: Tesseract (0.7720) beats the Migne-trained primary engine (0.6928) by +7.92pp on Greek body, which is why both engines' outputs are shipped in ocr_challenger. Neither engine output is a transcription.
• Not claims-ready. No APOC volume is claims-ready, and no downstream claims consumer exists yet for this corpus. Nothing in this dataset should be cited as a claims-bearing text. The 60 residual Latin rows in APOC_002/003 that earlier disclosures described as awaiting disposition now carry terminal dispositions -- 21 retained with no candidate correction, 32 retained with an unsafe candidate correction, 7 retained pending reconstruction -- but this draft disposition is not folded into the shipped repair_sidecar config and stays outside this repo. The claims-universe definition used for these counts covers exactly 1,592 regions on the v6 accounting and 2,265 on the v7 accounting. A binding-closure draft (1,592 rows) exists internally, but all three volumes remain only partially source-bound on both accountings -- not a fully claims-ready universe.
• Apparatus is zoned, not decomposed as raw text. Critical apparatus is carried as raw text in ocr (Apparatus_R1-R3). An apparatus-entry grammar/parse probe on a 72-region sample (24/volume, 1,027 entries) read at a 72.05% parse rate. A later, fuller measurement (8,793 apparatus entries probed on 432 pages: 198 APOC_003 + 74 APOC_001 + 160 APOC_002) found a strict link rate (exact/normalised lemma match to a merged body line) of 2.45% overall = APOC_001 6.12% / APOC_002 4.69% / APOC_003 1.14% (Lipsius-Bonnet is the worst-linked volume: its line-number keys are merged into body lines by the segmenter); an extended heuristic tier reaches 13.62% but carries a demonstrated false-positive rate and should not be used for claims. The blockers are OCR-side: superscript line-number keys are lost on the majority of candidates. Superseded 2026-08-24: the apparatus IS now decomposed -- see the apparatus_entries config and its own section below. The linkage figures in this bullet are unchanged (still near-zero, still OCR-blocked) and stay accurate for apparatus-to-body-line linkage, which apparatus_entries does not fix.
• APOC_003 (Lipsius-Bonnet) scan bleed. Resolved at the v4 re-OCR: gutter bleed from the source scan is zoned as Marginalia_ScanBleed (7,264 APOC_003 rows) and marginal line numbers as Marginalia_LineNumber (1,804) — both non-body labels, excluded from works by construction. An earlier disclosure of scan bleed affecting 71.7% of a Marginalia-labelled sample is historical and resolved by this re-zoning.
• Region labels are zoning heuristics, not ground truth. Some witness-sigla blocks are labelled MainText_ColLatin; some Apparatus_R3 regions on index pages are index entries. Page-level classes (title page, imprint, conspectus siglorum) are now classified (2,027 rows, one per volume/page, 0 unresolved, 12-value enum) and six conspectus pages (all APOC_003, the two 'Notarum explicatio' tables) have been corrected, but neither table is folded into the ocr/works configs shipped here -- they remain candidate tables outside this repo. A frontmatter/apparatus-boundary correction (1,705 rows) also stays a proposal without an acceptor.
• Sigla table is candidate-only. A hand-transcribed witness-sigla table exists outside this repo and is not claims-grade. It now covers 319 sigla across all 39 works (310 high confidence; the missing work's list was located via a 300-dpi text-layer scan of the Lipsius prolegomena), still hand-transcribed and still not shipped in this repo.
• works boundaries. An earlier accounting pinned 39 works / 1,592 regions and flagged 8 Lipsius-Bonnet Pars II works whose documented page_start excluded their own title/sigla head page. The current work_spans_v2 spans fix both: the Pars II head leaves are now pulled into their spans, and APOC_001 (previously only 3 of 20 works spanned, leaving ~332 body pages with no work_id) is now covered in full by the printed Tischendorf conspectus (see the table above). works now ships 56 rows (20/11/25) over 2,237 regions -- see the works config row above.
• RESOLVED (v13): the 41st repair now surfaces in works. Through v12 the martyrium_andreae_alterum repair (APOC_003 p534 r15) was fully bound but excluded from the work text because the raw region was labelled MainText_ColLatin while the span selects ColGreek. v13 shipped the reviewed_relabel (structural_role_sidecar row 4,061) after render adjudication, and the corrected paragraph (§2.3) is now IN the work text (+126 chars). The original caveat is preserved in the git history of this file as the audit trail.
• Provenance. Raw ocr parquets are byte-immutable (re-hashed at every revision; hashes unchanged across the v4 re-OCR); all corrections ship as additive sidecars keyed by source_unit_id. Sources are public-domain 1866-1903 editions; pdf rows are pointers into the source archive, not copies.

What may be stated without caveat: three editions, 22,575 ocr rows (APOC_001 2,957 / APOC_002 1,036 / APOC_003 18,582, v4 re-OCR zoning), 24,664 ocr_challenger rows, 56 works rows; 41 reviewer-accepted additive text repairs + 78 structural-role corrections + a 1,019-row FP-receipted confusion sidecar + 27,049 decomposed apparatus entries, shipped as sidecars; raw OCR byte-immutable across every assurance and parity check to date.

apparatus_entries (added 2026-08-24)

One row per decomposed critical-apparatus entry (lemma + variant reading(s) + witness sigla), both editions combined: APOC_001+APOC_002 (Tischendorf, 7,010 entries over 748 regions) and APOC_003 (Lipsius-Bonnet, 20,039 entries over 1,008 regions) = 27,049 rows. Each row is keyed on the same composite anchor as ocr -- (volume_id, page, anchor) plus entry_idx for the entry's position within its region -- not the verse address plane: apocrypha apparatus notes a manuscript witness disagreement, not a scripture reference, so there is nothing for osisRef to resolve against. osisRef and match_grain='apparatus_entry' are still present on every row (granularity HARD RULE); osisRef is null throughout, same convention as ocr/works/confusion_sidecar -- apparatus entries have no scripture address to resolve against, by design.

Both editions: edition_family is tischendorf or lipsius_bonnet; the two are read by separate sigla lexica and were run through the edition-separation firewall as one combined population (first time both editions were firewalled together) -- 11 of 11 checks PASS, 0 failed; the unfirewalled false-resolution rate would have been 16.77% (11,457 of 68,304 sigla tokens would have falsely resolved against the wrong edition's table).

Parse rate is disclosed, not hidden, and is lower for the busier edition: 65.08% (tischendorf, 4,564/7,010) vs 75.79% (lipsius_bonnet, 15,192/20,039) parse as a clean lemma/variant/sigla structure (form != unparsed); the rest keep their raw entry_text with form='unparsed', lemma_text/variant_text null-or-partial, and are not silently dropped. Passing parse_ok=true is a structural parse (the sigla-notation grammar matched), not a content guarantee: a minority of parse_ok=true rows are long multi-witness variant quotations that read as continuous prose rather than a terse critical note -- e.g. vindicta_salvatoris (APOC_001 pp.573-576) carries lemma_colon/lemma_cum entries of full parallel-witness sentences (Vercelli/Ambrose/Anglo-Saxon recensions quoted in line) up to the 400-char entry_text cap. The sigla/witness fields on these rows are still structurally correct; only lemma_text/variant_text read as prose rather than a compressed apparatus note. No lexicon or length filter was applied to reclassify or exclude them -- they are shipped as the engine parsed them, same disclosure standard as every other caveat in this card. On an n=15 stratified semantic sample of parse_ok=true rows (seed 20260824b, independently re-checked 6/6 at the promotion gate), 46.7% were fully correct lemma/variant splits, 26.7% mis-split, 26.7% not-apparatus content -- read parse_ok as parse-shape, never as a correctness rate.

21-row residual, excluded, not silently dropped: of the 1,777 true-population apparatus regions (v9 raw 3,143 minus the 1,366-row index/TOC structural-role exclusion), 21 -- all APOC_003 Apparatus_R3 -- have null/empty OCR text and are dropped by the (unedited) apparatus5_full_decomposition.py population filter before entry decomposition even runs; this is why regions_actually_decomposed (1,756) is 21 short of true_population_rows (1,777). It is an OCR-quality gap in those 21 regions, not an apparatus-decomposition defect, and is tracked separately from the index/TOC exclusion in the lane receipt.

Lineage: the apparatus-decomposition output was promoted unchanged into this config's publish copy with three added promotion-metadata columns (candidate_only=false, publication_authorized=true, promoted_date='2026-08-24'); every source column is kept, none dropped. One open item is not resolved by this promotion: apparatus_linkage (entry <-> body-line binding) still emits effectively 0 edges (1 of 27,049) -- decomposition and linkage are separate capabilities, and this config ships the former only.

Revision v9 (2026-08-23)

Raw OCR is the v4 re-run (zoning v4; per-volume engines as above), promoted after a 7-receipt validation against the prior stage (raw lineage to the compute jobs, works reproduction, region-level drift root cause with image adjudication, sidecar survival 84/84, confusion register). repair_sidecar (41 rows) carries the reviewer-accepted repairs re-anchored onto the v4 regions plus one two-reviewer-accepted region re-OCR; structural_role_sidecar (78 rows) excludes page-number stamps from the derived claims feeder and restores one apparatus-mislabelled paragraph; confusion_sidecar (1,019 rows) is the lexicon-gated, ink-FP-receipted κ/χ correction layer for APOC_003 (raw never mutated; a declared, owner-accepted ω-family residual remains display-grade only — see the confusion register in the pipeline receipts).

Revision v15 (2026-08-29)

The declared κ/χ residual is CLOSED. All 29 loci flagged at v14 were adjudicated per locus with a reproducible measurement (per-glyph ink-below-baseline decomposition, calibrated on known κ and χ in the same lines) plus independent eyes-on: every one of the 29 measures κ — 29 normalized, 0 kept as printed χ, 0 residual. The v14 sentence describing ≈ 29 unnormalized loci is superseded. All ops are single-character χ→κ substitutions: exactly 3 works' text_sha256 changed, 0 characters of length anywhere. Honest counter-evidence, recorded rather than hidden: at one locus (κατάκρισις, acta_thomae) the raw OCR and the shape-based confusion sidecar both read χ; the lexicon is neutral; the ink is unambiguous (0px below baseline, identical letterform to the word's own initial κ, while the adjacent ρ descends 14px — the earlier χ reading had measured the neighbouring letter). Also corrected this revision: nothing else — raw ocr/ocr_challenger/pdf and both role/confusion sidecars are byte-identical to v14.

Revision v14 (2026-08-28)

Completion of the per-row adjudication: the held acta_thaddaei repair and the extended defect lists for martyrium_petri_graecum and acta_thomae, every token re-confirmed on full-aspect renders (pdftoppm -scale-to-x 2500 -scale-to-y -1 -- the earlier convention omitted -scale-to-y and produced 2:1 vertically squashed images on exactly the axis accents and descenders live on). acta_thaddaei: 14 narrative tokens corrected while keeping the candidate's broad apparatus recovery (incl. Ἰουδ., ἐπιβ., θερα-πεύσει, σοὶ with grave, apparatus marker ¹³, βουλό-, sect. 3 l. 13). martyrium_petri_graecum: 16 ops incl. witness sigla cum AS / A restored (a wrong siglum is a content-altering defect), the dropped pipe restored, lemma ἢ, μακάριος, ἦσαν, ubi, πνεῦμα. acta_thomae: 7 ops incl. οἰκτίρ- (kappa), πρᾳότητος (subscript confirmed in print), the truncated cross-reference footnote restored to 'cf Luc 6, 36' and re-ordered after column B (footnotes are debris-dropped from works, so this changes the sidecar record only). Exactly 3 of 56 works changed (+13 chars corpus); 53 byte-identical to v13. Typeface ruling (this revision): in this edition χ carries a below-baseline descender and κ does not -- the two are separable in print, so kappa/chi is a measurable axis, not a free variant. The shipped repair texts still contain ~29 κ↔χ loci transcribed as χ by the v7-era re-OCR where the page sets κ; these are DECLARED here, not yet normalized -- each will be corrected only against its own render (the ἀποστύλου precedent: a rule-based sweep would overwrite genuine printed forms). ἀποστύλου itself was re-confirmed a FOURTH time at 6x zoom against same-line omicron controls -- it is what the page prints.

Revision v13 (2026-08-28)

Per-row, render-adjudicated resolution of the v12 card's 7-repair caveat (owner directive: adjudicate per-row; every verdict below rests on a fresh page render and its own transcription, with three disputed readings settled by an independent third render). Outcomes: 4 repairs corrected in place (acta_ioannis x2: tj→τῇ, xvolov→κυρίου, ser→scr; martyrium_petri_graecum: xvglo→κυρίῳ; acta_thomae: vm→ὑπʼ, DO→ϑεῷ) — the proposed iota subscript on ἐπιούση was checked against the print and REJECTED (the page shows it unsubscripted); 2 repairs WITHDRAWN (acta_andreae_et_matthiae p542, acta_philippi p861: the re-OCR candidates wreck near-verbatim narrative; their regions revert to raw). Withdrawn rows are RETAINED in repair_sidecar under status=withdrawn_after_render_adjudication — row count stays 41; a withdrawn row never binds into works. 1 repair still under adjudication (acta_thaddaei p391: genuinely mixed — broad apparatus recovery against real narrative degradation; its text is unchanged from v12 pending a dedicated fix set). 1 region relabel (structural_role_sidecar 4,060→4,061): APOC_003 p534 r15, three Greek body lines of martyrium_andreae_alterum mislabelled MainText_ColLatin, now MainText_ColGreek — the work gains its §2.3 paragraph (+126 chars), with the repair candidate's print-contradicted tokens fixed first (2,8→2.3, δύμιοι→δήμιοι, χρατοῦγτες→κρατοῦντες, τύπον→τόπον; ἀποστύλου kept — it is what the page prints). Exactly 6 of 56 works changed; the other 50 are byte-identical to v12 (poison-controlled). Validation: boundaries_validate_apoc_successor_stage_v2 PASS with the new unchanged_span_text_revised_by_sidecar class (raw OCR byte-identical; works reproduction gate proves every change is sidecar-caused). Known raw defect left standing: a neighbouring region of martyrium_andreae_alterum prints τόπον but raw OCR reads τύπον (pre-existing; raw is immutable; queued for the next repair batch). The verifier-noted extra defect lists for the corrected rows (7 + 4 tokens) are deferred to the next batch pending re-shoots at the estate render convention.

Revision v12 (2026-08-28)

Single-token fix-forward closing the defect the v11 correction below discloses. Repair row apoc_repair_9c0d2de7 (acta_thomae, APOC_003:p0982:r006) had its candidate_after_text token lov-δαν corrected to Ἰού-δαν -- the row's own before_text and the raw OCR both already read Ἰού-\nδαν; the re-OCR candidate had latinized the first three letters. works.parquet was rebuilt: exactly ONE work changed (acta_thomae), and the entire corpus diff is that token (v11_text.replace('lovδαν','Ἰούδαν') == v12_text, verified). Raw ocr/ocr_challenger/pdf remain byte-identical to v10/v11. The v11 sentence claiming the token could not be located stands withdrawn below as audit trail. On the 7-repair caveat below: Latin-letter runs are NOT automatically defects -- the Lipsius-Bonnet critical apparatus is printed in Latin, and at least one of the 7 (martyrium_petri_graecum) was independently render-confirmed as a genuine improvement (legible Latin apparatus replacing Greek-glyph mojibake). Per-row adjudication of the 7 is owner-gated and pending.

Revision v11 (2026-08-28)

APOC convergence, three owner-accepted corrections (authorization 2026-08-24T01:25:02Z, quoted below). Raw ocr/ocr_challenger/pdf and works.parquet are BYTE-IDENTICAL to v10 (works_sha256 unchanged, re-verified by an independent boundaries_validate_apoc_successor_stage_v2.py --allow-raw-ocr-revision PASS). Only repair_sidecar.parquet changed: the one row with reanchor_status=null (see the works caveat above) had its reanchor_* columns filled -- still 41 rows.

CORRECTION to the v11 card as first published (2026-08-28, integrator): the sentence published at v11 saying the reported 'lovdan' mojibake token "was searched for exhaustively across every text-bearing column and could not be located" is FALSE and is withdrawn. The search tested the ASCII string lovdan; the actual token is MIXED-SCRIPT -- lov (Latin l/o/v) + δαν (Greek) -- so an ASCII query could not match it and the zero result was guaranteed by construction, not evidence of absence. The token is present in this revision's works.parquet (acta_thomae, in both text and text_normalized): the published reading is ἥτις lovδαν ᾿Ισκαριώτην where the raw OCR reads Ἰού-\nδαν (Judas Iscariot).

Systematic caveat -- an accepted repair can make the text WORSE: 7 of the 41 rows in repair_sidecar replace raw regions containing ZERO Latin-letter runs with candidate text containing 4-163 of them (379 introduced in total). All 7 are status='reanchored_v4', language='grc', volume APOC_003, and all 7 bind exactly, so all 7 ARE substituted into works.parquet: acta_ioannis (x2), acta_andreae_et_matthiae, martyrium_petri_graecum, acta_thaddaei, acta_thomae, acta_philippi. Root cause is recorded in the sidecar's own reanchor_basis: these are v7-era reviewer-accepted corrections carried forward onto raw that has since been re-OCR'd BETTER. Each of the 7 records its own staleness verbatim -- similarity(SequenceMatcher.ratio) of the accepted v7 text against current raw is 0.269 / 0.604 / 0.647 / 0.680 / 0.688 / 0.698 / 0.875, every one below the 0.95 bar, each annotated (<0.95 => v8 raw does not reproduce the reviewer-accepted v7 correction; carried forward unaltered). Read 41 reviewer-accepted text repairs with this qualification: acceptance was against the v7-era raw, not against current raw. Remediation is owner-gated and not applied here.

Publication-authorization bookkeeping was corrected in this stage's PUBLICATION_AUTHORIZATION.json (owner quote: "recommendations accepted for 1-4, abide by user goals and intentions for the apoc task"); claims_ready stays false (PC-C7 remains owner-gated).

Gated dataset — request access via this page. License: public domain (three pre-1929 imprints); the OCR derivations are released under the same terms.