Elamur Groupantsenglish-orchestration
The groupant beside models/english-orchestration.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 |
context-transport | representation_transport | partial source derived model | context-transport |
effect-prefix | effect_boundary | partial source derived model | effect-prefix |
coordinate-join | coordinate_system | partial source derived model | coordinate-join |
uncertainty-join | assurance_boundary | partial source derived model | uncertainty-join |
Model text
descriptant
english-orchestration: wave27
main imports en modules declared by lib; license gate is only on train dispatcher, not a general evaluation gate. ctxeval consumes ctx::gold and ctx::Rng plus ud::Scores. diag calls parse baseline/kbest/constrained and uses gram enums. emode force/fix connect to annotation/tagger and embedded registry/dictionary; build output binary indices must match dict consumers (not independently validated here). Shape agreement by ID/token count does not prove form identity or valid trees.
This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave27 report. Registration is not a universal proof or whole-system runtime claim.
model
english-orchestration
forward
CLI selected command composes argument projection with library calls. train: commercial_gate(paths) then output argument check then delegated train/save/reload; no train call on rejected license. Context evaluation: Setup→gold/derived/shuffled→masked Ctx→tagger/parser features→aligned System runs→seed means and paired bootstrap. build.rs produces little-endian lexicon/layout/constants consumed by dict/morph and thus EModE ending decisions. diag consumes parser outputs and gold geometry, with constrained decode failures falling back to baseline.
invariant
Module declarations establish one shared namespace but prove no module semantics. Context mask removes layers, coordinates remain zero-based token positions versus one-based heads. License admission reads path names only. Paired bootstrap samples same sentence indices for both systems. No global transactional guarantee: file/stream prefixes survive later failures.
cost
Add selected adapter/delegated costs; no sum over unselected commands. build indexes expected linear/worst quadratic; ctx stype O(n²), mention handling O(bytes+mentionsstack depth), shuffle has unbounded rejection-loop worst case. ctxeval O(SRT+SRBN+SB log B), diag geometry up to O(n³) plus decoder cost. CLI folds O(kbytes). No speed measurement or claim.
composition
main imports en modules declared by lib; license gate is only on train dispatcher, not a general evaluation gate. ctxeval consumes ctx::gold and ctx::Rng plus ud::Scores. diag calls parse baseline/kbest/constrained and uses gram enums. emode force/fix connect to annotation/tagger and embedded registry/dictionary; build output binary indices must match dict consumers (not independently validated here). Shape agreement by ID/token count does not prove form identity or valid trees.
domain
Bounded synthetic strings, valid aligned token arrays and in-range heads for geometry; valid System shapes with >=1 baseline and seeds, bootstrap B>=1 for comparison. Effects limited in tests to isolated temporary synthetic build/fold/report files. Real training/model/corpus paths are source-derived only.
limitations
No corpus, teacher, API, GPU, held-out or training run. Existing compiled dictionary/registry tables participate only where safe tests call ordinary functions; dictionary-sensitive early-return tests excluded. u32 narrowing, allocation failure, arbitrary malformed heads and environment OnceLock alternate modes remain gaps.
checks
- 14 named Rust tests twice debug and twice release, identical sorted outcomes, exits0
- Build entrypoint empty/synthetic/tag-before-id rejection scenarios twice per profile; successful outputs byte-identical
- 7 actual compiled mutations, compile0 / assertion101 / exact restore / rerun0
- 8 paired syntax projections identical; 5 source-scoped main fragments
- Strict Clippy exit101 on existing build.rs:51 manual_is_multiple_of
catalog
behavior-complement
Across train dispatch, exclude model training/saving when any path first fails commercial_gate; across folds k<2 exclude input read and directory writes; across test_ctx None exclude path reads. Within aligned reports exclude independent baseline/treatment bootstrap draws. These scoped exclusions do not forbid partial output after later I/O failures.
reverse-reconstruction
Given accepted gate and rendered report, compatible antecedents are all path lists mapping to admitted sources and aligned gold/system runs producing the rounded means/intervals/group counts. They include arbitrary directory spellings, different per-seed outcomes with equal aggregates, erased annotation features and omitted examples. No unique reconstruction from report/gate result; feature ablations and constant/hash collisions further merge antecedents.
context-transport
Doc nonempty sentence order→Ctx/docs; train_ctx filters tagged sentences and offsets document IDs; test_ctx validates pass1 IDs/counts; masking occurs after derivation/shuffle. ctxeval System loader checks IDs/counts but not forms. Crossing these adapters requires explicit order alignment, not just equal lengths.
effect-prefix
Build reads then writes lexicon/layout/words sequentially; folds creates dir then fold/rest files; train saves before reload comparison; conversion streams completed files before later failure. A failure may retain prefixes and no rollback is modeled.
coordinate-join
Token vector index i maps head ID i+1; zero is root. Context entity features use zero/no-layer, one/outside, >=2 encoded type/status. Build offsets distinguish byte offsets from record counts and symbol IDs; none are interchangeable.
uncertainty-join
Compile visibility of lib is not execution evidence. Fourteen finite tests and seven mutants falsify selected contracts only; unknown live paths and unsupported percentage approximants remain explicit. Bootstrap intervals are sampling summaries, not proofs of real-task improvement.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/english-orchestration.groupant/