README.md

---
license: other
license_name: mixed-per-partition
tags:
- typology
- multilingual
- bible
---

# gram-struct — per-language grammatical fact sheets

One JSON per language, fused from every source the lexeme-aligner already has,
split into four provenance partitions so that licensing is a directory boundary. Design and contract:
`docs/architecture.md` §2. Built by `python3 -m lexeme_aligner.gram_struct --build`; never hand-edited.

| partition | license | contents | provenance |
|---|---|---|---|
| `external/` | **CC-BY-4.0** | direction slots + existence facts | Grambank v1.0 (Skirgård et al. 2023, CC-BY-4.0; Glottolog for the Glottocode→ISO mapping, CC-BY-4.0) and WALS (Dryer & Haspelmath 2013, CLDF v2020.5, CC-BY-4.0). Every fact cites its source codes (`GB074`, `WALS:85A`). |
| `imputed/` | **CC-BY-SA-4.0** | direction slots only | lang2vec / URIEL `syntax_knn` (Littell et al. 2017, CC-BY-SA-4.0) and URIEL+ (Khan et al. 2025, CC-BY-SA-4.0; `adposition`/`object_verb` only, the slots that cleared 90% agreement with Grambank) — kept physically apart because share-alike applies to derivatives of these values. |
| `derived/` | **CC0-1.0** | our own alignment statistics: direction slots `adposition` (+ `adposition_word`: is the adposition a separate word), `possessor`, `subject_verb`, `object_verb`, each with per-edition votes (`editions`, `agreement`); `article_bound` (is the definite article fused into the noun, see `article_bound.py`); the constituent-order profile; `audit.multiword_rates` | computed from this project's eflomal(+gloss) alignments of every edition of the language, MACULA syntax only (no BHSA-derived field); no source text redistributed. Agreement with Grambank where both exist (2026-10-08): adposition 92.6% (n=390), possessor 91.7% (n=192), subject_verb 88.6% (n=332 — inside the 85-90% `experimental` band of the project's admission rule; no alignment mechanism reads it), object_verb 93.5% (n=92); with Östling & Kurfalı's projected word order (2023): adposition 93.9% (n=212), subject_verb 100% (n=711), object_verb 99.5% (n=202) — a correlated method (both project from Bible alignments), reported as a consistency check, not independent proof. |
| `measured/` | **CC0-1.0** | mechanism verdicts, dated, gold named | human-recorded after scoring against gold; never inferred. |
| `kin/` | **CC-BY-4.0** | direction slots only, gap-filling | Glottolog genetic relatedness (bcv-query's `languages.db`) — the nearest external/imputed-resolved relative's value, lowest priority, structurally unable to override anything else; confidence is a leave-one-out band rate, not the individual fact. |
| `<iso>.json` | mixed (see above) | the merge of the five | convenience view; identical content, one file. |

## Contract

- A slot with `"direction": null` **and a `source`** is a fact: the source knows the codes and they
  resolve to no single side (e.g. Hindi `article`: GB022 = GB023 = 0). An **absent key means unknown**:
  no source covers this language for this slot. Consumers must gate on the difference.
- Existence facts carry `flagged` (the polarity's own meaning, from `analyze_language.RISK_RULES`) and
  `present` (the plain reading), plus `polarity`, `codes`, `known`, `matched`.
- `derived/` changes whenever the alignments are regenerated and carries `content_sha256`;
  `measured/` changes only when a human re-measures and re-dates an entry.
- No field in any file is computed from another field in the same file.
- Merge priority per slot: external > imputed > derived > measured > kin. `_coverage.json` → `merged_slot_sources` says, per slot,
  which partition (or which external source) the merged view took, and how many of those values are a null-with-source.
- Not derivable from alignments, so not here: numeral classifiers, evidentials, tone/phonology, whether a bound form is an affix
  or a clitic written together (an alignment sees tokens; "fused" is the most it can say), reduplication.