Elamur Groupantshebrew-corpus-adapters
The groupant beside models/hebrew-corpus-adapters.groupant in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial | forward |
| ast | not materialized | |
| composition | partial | composition |
| cost | symbolic source derived | cost |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | see source local evidence | checks |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| id | type | label | where |
|---|---|---|---|
behavior-complement | behavior_complement | partial source derived model | behavior-complement |
reverse-reconstruction | reverse_reconstruction | partial source derived model | reverse-reconstruction |
byte-coordinate | representation_transport | partial source derived model | byte-coordinate |
dependency-coordinate | dependency_transport | partial source derived model | dependency-coordinate |
book-code | tagged_order_transport | partial source derived model | book-code |
speech-carry | state_transition | partial source derived model | speech-carry |
effect-boundary | partial_effect_trace | partial source derived model | effect-boundary |
Model text
descriptant
hebrew-corpus-adapters: wave31
Real joins: cue form/name->features->count/logp columns; tokenize byte spans->segments->talmud slices; MACULA Node0-based -> head1-based -> surface permutation -> emitted-row head mapping; macula html_text->tanakh NFC edition; read_sentence open_after->next open_before within book. Counterexamples: copying MACULA ordinal to Tanakh order misaligns books; treating Segment.end as talmud slice end loses punctuation; union of component complements is not complement of composed relation. L/M are tagged siblings only.
This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave31 report. Registration is not a universal proof or whole-system runtime claim.
model
hebrew-corpus-adapters
forward
Tagged product L+M: L performs tokenize->features(cues)->Counts/Model->Segmenter, corpus::clean->talmud::original and Segmenter->talmud::segments. M performs XML->Sent->tree->check->surface_order->conllu; main shares library consonants/keys and dispatches prag/tanakh. Prag consumes supplied10-column rows plus prior speech state. Tanakh edition composes library html_text with NFC. No runtime L->M call asserted.
invariant
Preserve UTF-8 byte coordinates in langseg until explicit text copying, column order HEB,1-based MACULA head transport through both permutations/row expansion, sparse1-based verse coordinates and tagged book-code identity. Both model families are lossy: labels/text cannot recover erased markup/groupings. Valid nonempty row/tree, finite signed scores and finite nonnegative switch costs are explicit.
cost
Add only reached sibling paths; do not sum independent commands or overlapping fragments as serial stages. Langseg tokenizer O(B²) worst-case unmatched '<', ordinary bounded markup O(B), scoring O(F log V), Viterbi O(nK),K8. Corpus cleaner bounded60 lookahead O(B), registry quota can rescan O(rows*E). Macula repeated scans/ancestor walks O(n²) and depth-dependent subtree traversal; joins O(R log V), output/storage proportional to bytes. No end-to-end timing or speedup inferred.
composition
Real joins: cue form/name->features->count/logp columns; tokenize byte spans->segments->talmud slices; MACULA Node0-based -> head1-based -> surface permutation -> emitted-row head mapping; macula html_text->tanakh NFC edition; read_sentence open_after->next open_before within book. Counterexamples: copying MACULA ordinal to Tanakh order misaligns books; treating Segment.end as talmud slice end loses punctuation; union of component complements is not complement of composed relation. L/M are tagged siblings only.
domain
Small synthetic strings/XML/arrays and tempfile model roundtrip under dedicated temp root. Scoring requires representable counts, column bounds, finite alpha>0 and nonnegative switch costs for objective model; rows nonempty10 fields, heads valid for tree/CoNLL-U helpers. Raw adversarial input outside those premises is not certified.
limitations
Source inspection covers CLI effects, but actual live corpus, teacher, frozen assessments, production DB/model data, downloads and lesson workflows excluded. No parser completeness, language accuracy or universal safety inferred from finite checks. Existing audit-before-guard and transitive-invalid-head caveats retained.
checks
- 30 exact native test names twice debug/twice release with identical sorted outcomes; native-commands.json
- 6 compiled source-behavior mutants red, exact byte/hash restore, focused green; mutation-results.json
- 10 syntax projections generated twice and byte compared
- 10 source leaves,10 fragments,120 typed views with explicit cost components
catalog
behavior-complement
Under admitted states, L cannot emit neutral-only tokens or invalid path columns, M cannot choose Aramaic on majority tie, PD edition cannot admit wrong licence. The sibling universe is tagged transitions; neither path implies execution/effects of the other. These are partial source-derived exclusions, not exhaustive complements.
reverse-reconstruction
Reverse each actual join by retaining compatible intermediates: segment label fibers of feature/count/score states; CoNLL-U fibers of group trees and permutations; cleaned verse fibers of markup and sparse input coordinates. Keep tag L/M and book code; dropped metadata and source order are ambiguous. Unique inverse only when original permutation/data retained.
byte-coordinate
Token boundaries are UTF-8 source offsets; Segmenter spans last token end but talmud slices to next start/text length. Punctuation stays preceding segment; no leading-neutral recovery law.
dependency-coordinate
Old word head h maps0->0 else p(h-1)+1, then emitted-row head first[h-1]+1. Valid heads and shared word sequence are required; interpreting Node0-based ids as heads is invalid.
book-code
MACULA library and Tanakh order differ, so cross-reader identity is book code plus chapter/verse, not integer ordinal. Missing/invalid coordinates map0 and unknown code39, creating noninjective fibers.
speech-carry
Prag read_sentence(row,open,rules) produces next open; quote recognition precedes closure/carry/bare. Book boundary resets externally. Langseg has no caller relationship to this speech state.
effect-boundary
Model save truncates before rows; CLI talmud writes original/English before optional map read; converter writes accumulated books/maps separately. No atomic rollback or no-read claim for audit guard. These effect traces are source-modeled, not executed.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/hebrew-corpus-adapters.groupant/