full-alignments-manual — hand-made word alignments in the compact container format

Every word of a verse, linked to the Hebrew or Greek source word it translates, as made by people (no statistics, no LLM): one layer per published gold source and edition. Function words (articles, prepositions, pronoun suffixes) are included, and so are the cases where a source word has no target word of its own (markers) or a target word was supplied by the editor.

The files use the format of compact-alignments (same paths, same _index/, same tokenizer), so a compact reader works on them unchanged; the extra channels are described below. The statistical full alignment of the same editions (every row of every method) lives in compact-alignments-meta, merged into each book's .meta.json and decoded with that edition's _layer.json.

Layers

A layer id is <edition>+manual+<source>, where <edition> is the compact-alignments edition id of the same text, e.g. e/eng/eng_BSB+manual+clear/ and e/eng/eng_BSB+manual+bsb-tables/. The list of layers, with rows and books, is in manifest.json.

source	what	licence
clear	Clear Bible alignments (github.com/Clear-Bible/Alignments): Hebrew OT by word id, Greek NT by Strong's number and occurrence	CC BY 4.0
helfi	HELFI Finnish (University of Helsinki), morpheme-level	CC BY 4.0
sword	SWORD ChiUns (Chinese Union Version)	Public Domain
bsb-tables	Berean Standard Bible translation tables (the publisher's own interlinear): markers, joint phrases, supplied and inflection words	CC0

Each layer's _layer.json repeats its source, licence and base text. Left out on purpose: the Clear Russian Synodal gold (positionally wrong in the source itself), the Clear Portuguese layer (machine-projected by Clear, not hand-made), and gold whose licence does not allow redistribution. Door43 (CC BY-SA 4.0) layers are in the separate full-alignments-manual-sa.

File format

Every layer is a folder <iso[0]>/<iso>/<layer id>/ with per-book files that look exactly like compact-alignments files and add channels for what compact does not carry. No file holds a word of any text: every target reference is a token position in the edition's own verse text (tokenized with compact-alignments' tokenize.js), every source reference an ordinal into compact-alignments' shared _index/ files.

<BOOK>_<hash>.json       main: one string per spine verse, position-parallel to _index/<BOOK>_lexemes.json
                         "srcOrd:span ..." = the winning link of each CONTENT source token (compact's own meaning)
<BOOK>_<hash>.meta.json  channels (a verse's entry is "" when empty; null = the channel has nothing in this book)
_layer.json              attribution constants and the profile table (char -> method / score / prior / extra / attribution)

<hash> = the last 5 hex characters of the book's content hash (compact-alignments book_content_hash), so a file only matches the exact text it was built on.

Channels (in .meta.json):

channel	content
fn	"fnOrd:span ..." links of FUNCTION-WORD source tokens (articles, prepositions, suffixes …; _index/<BOOK>_fn.json lists them in order)
wp	one profile character per main entry; . = a view of a multi-source row whose row is in rows
fp	the same per fn entry
wx	per entry, space separated: _ or ;-joined parts: t<ints> the gold's own target-token numbers, S<id> a source id other than the token's spine key, P<ints> the profile's arguments. Main entries first, then fn entries
rows	"srcs:span:prof[:ext]..." every row that is not a claimed winner: losing alternatives, multi-token source attachments, unplaced rows, markers, joint rows
off	{ref: [row entries]} rows with no spine source token (e.g. a table row whose source word is not in the Nestle 1904 / WLC spine)

Slots: c<N> = content token srcOrd N · f<N> = function word fnOrd N · p<N> = token N of a verse folded into another verse's group (two source verses = one target verse).

Spans: 5, 5-7 contiguous, 5,9 scattered, L3,1,2 an out-of-order list, ~ unplaced (verse not mappable), ! explicitly empty, @6e / @-1u a MARKER: the source word has no target word of its own and sits after target word 6 (-1 = verse start); kind u = untranslated (BSB -), e = rendered elsewhere (BSB . . .).

Row extensions: k<i> index (into the srcs) of the keyed token · v joint with the following phrase (BSB vvv) · s<pos,..> target words the editor SUPPLIED (BSB [x], not in the span) · i<pos,..> inflection words made explicit (BSB {x}, in the span, flagged) · t<ints> target-token numbers · S<id> source id · T<strong> the row's own Strong's number when it differs from the token's · P<ints> the profile's arguments.

Profile arguments: a number inside a row's prior is per-row data (e.g. span extension's spanext_name_before:12 = the target position the rule appended). The profile table stores the prior with # in its place and the numbers travel with the row (P12), in order; put them back to get the original prior.

Vetoed rows (statistical layer only): a row whose profile extra carries "veto": "<check>" was judged wrong by a word-level check (a pronominal suffix aligned far from its noun, a prefixed preposition on the wrong side for the language). It is never claimed into fn and never shown as a view; it stays whole in rows.

Derived on decode, never stored: lexeme / Strong's / lemma / stem / surface / gloss (from the source spine via _index/) and the target words (from the positions and the edition text). No BHSA-derived field is carried.

Verse mapping

Spans are keyed by the source verse (Hebrew WLC numbering in the OT); read the target words from the edition's mapped verse (compact-alignments README, "Verse numbering"). Source and target verses are paired with the verse-mapping tables of TVTMS (STEPBible Data, Tyndale House, Cambridge — CC BY 4.0, https://github.com/STEPBible/STEPBible-Data).

License

Each layer keeps its source's licence (table above); the dataset as a whole is CC BY 4.0. Attribute the source named in the layer's _layer.json.