NuBerea/unified-grammar

Unified Distributional Grammar (Greek + Latin + Hebrew)

NuBerea/unified-grammar — the j-layer construction × function × source × era matrix, mirroring the shape of NuBerea/distributional-lexicon (lemma × sense × source × era) for grammar instead of lexicon: every grammatical claim scoped, counted, basis-carrying (see METRIC SEMANTICS below), and traceable to corpus instances, with traditional grammars (Smyth, Gesenius-Kautzsch-Cowley, Allen & Greenough) admitted only as witness prose, never as evidence. Full design: grammar/plans/DISTRIBUTIONAL_GRAMMAR_PLAN.v2.md.

This repo replaces the plan's placeholder name distributional-grammar — the owner-chosen repo name is unified-grammar throughout.

Config consolidation (2026-08-24, owner GO 2026-08-23): the per-language greek_* / latin_* / hebrew_* / nt_* / ot_* / lxx_* configs (30 total) were merged into 11 cross-language configs (construction_candidate, construction_gold, construction_judgment, construction_cell, construction_apparatus, grammar_claim, grammar_source_claim, parser_calibration, alignment_positions, projected_heads, latin_syntax_silver_merged), each carrying a lang (or, for latin_syntax_silver_merged, parser_id / calibration_version) column as the discriminator instead of a config-name prefix, built by scripts/build_consolidated.py. The 30 retired per-language config names are tombstoned (status: deprecated) in ontology/manifest/assets.yaml with successor pointers to the merged config; nothing was dropped -- see each merged config's manifest notes: for the exact per-source row counts and column-harmonization list (e.g. gold_definiteness, examples, calibration_version). construction_ontology, greek_syntax_silver, hebrew_gold_heads, hallmark_loci, and crosslang_construction_alignment were not part of the consolidation and are unchanged.

Namespaces

Configs are namespaced greek_ / latin_ / hebrew_, plus unprefixed crosslang_ for the one cross-language config — exactly NAMING_CONTRACT.md rule 1 from the lexicon family (a config that ships to another repo keeps that repo's own real name, never a lang-prefixed alias; construction_ontology itself is also unprefixed, since it is the shared language-neutral inventory both arms key against).

Config table

All 16 configs below are live (plus claim_adherence/claim_application, added 2026-09-12 by the CARD-PARITY-CHECK repair — 18 live total). Row counts are MEASURED against the published parquet footer metadata (ontology/manifest/row_count_check.py --repo unified-grammar, full 16-config re-run 2026-08-28: 16 pinned, 0 drift, 0 missing) — not carried forward from an earlier claim. Corrected 2026-08-30: construction_judgment, construction_cell and grammar_claim carry the 2026-08-30 S5-refresh publish's own footer counts, and hallmark_loci is corrected to its live value (the 67 this table showed until now was a stale card figure, never the published count — the 2026-08-28 pin sweep read 15 and reported 0 drift; the 2026-08-30 superscript-recovery publish then took it 15 -> 186). The per-language-prefix naming this table used to show (greek_construction_candidate, latin_construction_candidate, …) was the pre-consolidation plan; those names were never published. Every row here is the actual live config name; where a config is cross-language, the discriminator column that replaces the old name-prefix is given instead of a separate row per language (see "Config consolidation" above).

config	discriminator	status	k	row_count (2026-08-30)	builder
construction_ontology	shared, unprefixed	live	curated	20	scripts/build_ontology.py::build
construction_candidate	lang (grc/la/hbo)	live	computed	481,614	scripts/build_consolidated.py
construction_gold	lang	live	computed	281,215	scripts/build_consolidated.py
construction_judgment	lang	live	computed	3,492	scripts/build_consolidated.py
construction_cell	lang	live	computed	557	scripts/build_consolidated.py
construction_apparatus	lang	live	computed	444,257	scripts/build_consolidated.py
grammar_claim	lang	live	computed	452	scripts/build_consolidated.py + scripts/build_grammar_claim_s5.py
grammar_source_claim	lang	live	cc-by-sa-4.0	626	scripts/build_consolidated.py
parser_calibration	lang	live	computed	1,859	scripts/build_consolidated.py
alignment_positions	pair/target_corpus	live	cc-by-sa-4.0	576,312	scripts/build_consolidated.py
projected_heads	pair/target_corpus	live	computed	383,972	scripts/build_consolidated.py
latin_syntax_silver_merged	parser_id/calibration_version	live	computed	1,496,956	scripts/build_consolidated.py
greek_syntax_silver	greek-only, not consolidated	live	cc-by-sa-4.0	27,229,279	scripts/greek_silver/parse_silver.py
hebrew_gold_heads	hebrew-only, not consolidated	live	computed	305,517	scripts/latin/derive_hebrew_gold_heads.py
hallmark_loci	lang, not consolidated	live	computed	186	annotation/build_hallmark_loci.py
crosslang_construction_alignment	shared, unprefixed	live	computed	20	scripts/build_crosslang.py
claim_adherence	shared, unprefixed	live	computed	450	scripts/build_claim_adherence.py::build_adherence_table
claim_application	shared, unprefixed	live	computed	978,022	scripts/build_claim_application.py::build_application_table
feats_grammar_bridge	rail (pg/clsgrc/pl/ref/schol/apoc/pseud_grc/pseud_la)	live	computed	98,339,696 (pl rail only, 2 shards; see below)	scripts/build_feats_grammar_bridge.py::main

(The two rows above were added by the 2026-09-12 CARD-PARITY-CHECK repair — real live assets.yaml configs from the matcher-republish wave 2026-09-01 (B2/B4) that had never been declared in this card's configs: block; facts taken from the assets.yaml node only.)

candidate breakdown by lang on construction_candidate (sums to 481,614): grc 75,850 / hbo 205,365 / la 200,399.

feats_grammar_bridge (T12b, pl rail, live 2026-09-12)

Cross-rail token-grain bridge joining a token's real Case/Mood/Tense/Voice feature bundle (feats — the 5-tuple featslemma lemmata upstream, not the older upos-only 3-tuple) to an Allen & Greenough or Smyth grammar section_id/section_part/section_heading, one row per token (excl. PUNCT/SYM). rail discriminates the corpus each row came from (pg, clsgrc, pl, ref, schol, apoc, pseud_grc, pseud_la — all siblings of this same config, joined into the manifest's unified-grammar/feats_grammar_bridge node one rail at a time).

pl rail (Migne PL, Latin) — status: live in assets.yaml, published in two shards under data/feats_grammar_bridge/ (picked up together by the train-*.parquet glob above): train-00000-of-00001.parquet (full build, T12B_PL_BUILD_RECEIPT.md/ T12B_PL_PUBLISH_PREP_RECEIPT.md) + train-00001-of-00002.parquet (pl-03/pl-07 remainder, T12B_PL_INCREMENT_RECEIPT_v1.md):

• 98,339,696 rows total, 217 distinct source_volume — the PL rail's full, contiguous PL_001..PL_217 (0 gaps). Shard 0: 82,132,029 rows / 185 volumes (PL_001..PL_098 + PL_113..PL_199). Shard 1: 16,207,667 rows / 32 volumes (PL_099..PL_112 + PL_200..PL_217, the pl-03/pl-07 remainder that landed after shard 0 was built) — 0 source_volume overlap between the two shards, schema byte-identical.
• 24-column schema byte-identical to the live patristics-migne-pg/classical-greek feats_grammar_bridge siblings.
• osisRef empty-string on all rows (no verse-id coupling for this rail yet, same posture as patristics-migne-pl/lemmata); anchor/match_grain/volume_id/page/region_idx 0% NULL (shard 0 and shard 1 both verified).
• basis distribution (shard 0): upos_topic 35.96% / feat:Case=Nom 14.95% / feat:Mood=Ind 12.94% / feat:Case=Acc 11.46% / feat:Case=Abl 10.17% / feat:Case=Gen 6.30% / unmapped 3.88% / remaining feature-specific bases <3% each. Shard 1's distribution is proportionally consistent (upos_topic 35.43%, feat:Case=Nom 14.81%, ... unmapped 3.79%).
• section hit-rate (lat): shard 0 96.12% (78,946,375/82,132,029); shard 1 96.21% (15,593,449/16,207,667) — combined ~96.14%.
• k_firewall: ships A&G section_id/part/heading POINTERS + upos/feats VALUES + a basis/confidence label ONLY — no A&G prose_text copied or reconstructible (same firewall as pl_grammar_bridge; permissive consumers call ag_sections/mcp__nuberea__ag_sections with section_id for the actual grammar-author prose, ref-cc-by-sa-sharealike-scope owner ruling 2026-07-03).

Upstreams

Live, already-on-platform (no new intake needed for Phase 0): macula-greek, macula-sblgnt-syntax, macula-hebrew, macula-hebrew-syntax, septuagint-lxx, vulgate-clementina, features/grc_syntax, features/lat_syntax, classical-greek/grc_segments, distributional-lexicon (lexical-class covariate).

Planned t0 repos (owner-gated creation, per-repo confirmation — see plan v2 §2, §8): papygreek-treebank, agdt-treebank, pedalion-treebank (Greek gold treebanks, Phase 2), latin-llct-treebank, latin-udante-treebank, latin-circse-treebank (Latin gold treebanks, Phase 3), allen-greenough-grammar (Latin grammar witness, DCC re-edition, CC BY-SA 3.0, Phase 1), gesenius-kautzsch-grammar (Hebrew grammar witness, Wikisource, CC BY-SA 4.0, flagged incomplete, Phase 1).

smyth-grammar — UNBLOCKED 2026-08-23 (owner ruling): Smyth 1920 is public domain; the Perseus/ Alpheios CC BY-NC-SA claim attaches to their encoding, which is not redistributed. NuBerea/smyth-grammar (k=public-domain, 3,261 sections §1–§3048 + 213 dialect paragraphs, 665-row TOC; built from the Perseus-derived neelsmith/smyth TEI, Beta Code → Unicode) is the Greek grammar witness; greek_grammar_claim / greek_grammar_source_claim can now be implemented against it (plan v2 §8 D6 superseded).

k / licensing

• construction_ontology: k: curated — the owner-ratified 20-construction seed table, hand-authored, language-neutral, holding grammar_crosswalk section-NUMBER pointers only (never grammar prose).
• Every derived {lang}_* and crosslang_* config: k: computed once implemented — t = max(upstream t) + 1.
• {lang}_grammar_prose (not yet a config in this repo's CONFIGS_BY_LANG — ships from the individual grammar-witness t0 repos per plan v2 §3) carries the true source-license province via the license-home exception, mirroring greek_lexica_prose.

k_firewall (applies to every {lang}_parser_calibration config, present and future): "Ships construction ids, counts, scores, confidence and section-number pointers only — no treebank tokens or grammar prose redistributed or reconstructible." This mirrors the estate's edition_delta / greek_lexica_prose firewall pattern (samaritan-pentateuch-gall / dss-levy precedent): NC-licensed treebanks and benchmarks (Gorman, UD-Perseus, UD-PROIEL, GLAUx's NC layer, UD_Latin-ITTB/Perseus-LDT/PROIEL, LASLA, ETCBC BHSA, UD_Ancient_Hebrew-PTNK) are registered as NC-LICENSED t0 (never-wired) instruments — private, never tenant-wireable, never public or gated-public, never in any config's upstreams: list. Enforced by scripts/grammar_never_anchor_guard.py in build.py --verify: FAIL = no publish.

The witness guard

scripts/grammar_never_anchor_guard.py (transplant of distributional-lexicon's scripts/hebrew/never_anchor_guard.py, which keeps SDBH's CC BY-SA domains from leaking into the CC BY lexicon) asserts:

1. No *_construction_cell / *_construction_apparatus / *_grammar_claim parquet carries a section_id / grammar-work-key column, named or disguised.
2. Every *_grammar_source_claim row has evidence_status == 'grammar-author-claim'.
3. grammar_crosswalk pointers live ONLY on construction_ontology — no other config may carry that column.

Traditional-grammar prose is witness, never evidence: a grammar_explanation tool answer may cite that a construction's function reading is consistent with Smyth §2054 or GKC §116, but no published cell, apparatus, or claim row may be keyed on, or derived from, a section id.

NAMING_CONTRACT policy (plan v2 §4, condensed)

The estate's edition_delta (samaritan-pentateuch-gall / dss-levy) precedent applies verbatim to any future NC-benchmark instrument read in this repo: an exception: {class: instrument-read, reason: ...} block on the manifest node, a k_firewall string stating exactly what ships and asserting non-reconstructibility, k staying the shipping side's true class (never inherited from the NC upstream), and the NC source itself registered as a quarantined instrument — never wired, never an upstreams: entry, never public. construction_id naming is language-neutral by construction (SS5.1/SS5.4 lint in build_ontology.py): no GREEK/LATIN/HEBREW/GRC/LAT/HBO token may appear in a construction id, so the ontology scales across languages without ever encoding a language into the join key itself.

Hard rules honored by this repo

No Hugging Face operations, no git commit/push, no repo creation performed by any script here. No LLM calls inside any builder (stage-prompts-only rule) — the ontology seed is hand-curated, not model-generated. No embedding work, no OCR, anywhere in this repo's scope. assets.yaml and ontology/manifest/schema.py are read-only references for this repo, never edited by it.

METRIC SEMANTICS (2026-08-24)

There is no per-instance confidence in this dataset, and the column names say so. nb_posterior (apparatus) / mean_nb_posterior (cell) is a per-construction naive-Bayes posterior whose only features are genre + the extractor's variables; 16.4% of matched candidates carry no variables at all, so it is a construction+genre base rate — hbo GEN.NN takes 6 distinct values across its 54,260 silver rows (all layers together show 14, because proxy rows carry a re-pointed per-row posterior), hbo CL.PURP exactly one. predominance_share (claims) is a descriptive share of the model's OWN output distribution and is therefore circular as a truth signal.

The measured trust signal is holdout_agreement — leave-one-source-out top-1 agreement — now carried on apparatus and cell rows, not just claims. Read it with n_train_labels/n_train_classes: n_train_classes == 1 means a constant model whose posterior is 1.0 by construction (10 of 47 construction-language pairs), and 5 claims rest on a single training label. hbo INF.ART averages posterior 0.804 across 5,474 rows at a hold-out of 0.1.

On a layer='proxy' row the label came from the adjudicated panel, so a LOW nb_posterior marks model-vs-human disagreement, not doubt — those are the best-evidenced rows here.

Documented gap (2026-08-28): latin_syntax_silver_merged.parse_conf is entirely null across all 1,496,956 rows (verified against the published parquet: null_count == num_rows) — the silver parsers backing this config emit no per-token confidence, so the column carries a schema slot with no signal in it. Use parser_calibration (joined on parser_id / calibration_version) for the actual trust signal on Latin silver syntax; do not read parse_conf as a null-vs-low-confidence distinction. This is a known gap, not a bug — see Phase C item 4 in grammar/plans/ALIGNMENT_PLAN_2026-08-28.v1.md for the standing decision this still needs (surface the calibration join in tool descriptions, or drop the dead column repo-wide).