Elamur Groupantstranslation-evaluation-pipeline
The groupant beside models/translation-evaluation-pipeline.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 |
population-transport | representation_transport | partial source derived model | population-transport |
cost-denominators | metric_transport | partial source derived model | cost-denominators |
effect-order | failure_precedence | partial source derived model | effect-order |
coordinate-join | coordinate_transport | partial source derived model | coordinate-join |
Model text
descriptant
translation-evaluation-pipeline: wave32
Real joins: bible::read_pool/sources text→corpus PVerse; main run_sys→measure::Run→Agg/VRec→block_measure and report::write; many→corpus/measure plus own pair metrics; pages receives wlc_jps VRec stage arrays from main and Hebrew encoder outputs. many::agree skips punctuation while block_meta includes it: their counts are not interchangeable. Report groups collapse layers. No invented call from report to many.
This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave32 report. Registration is not a universal proof or whole-system runtime claim.
model
translation-evaluation-pipeline
forward
bible/sources supply VKey text maps; corpus groups and parses into PVerse. main evaluates target-population records with measure's reader/encoder/stack residuals and paired-source aggregates; blocks combines chapter-local layouts and reader scripts. many independently orchestrates multi-translation pair accounting from the same corpus/measure representation. report and pages consume residuals and normalized counts.
invariant
Explicit valid tree/key/population domains; residual-assisted exact target reconstruction remains distinct from normal translation scores. Original production bytes preserved. Finite arithmetic and representable counters required; Docs sample n>0; storage membership indices<64.
cost
Sum delegated parser/reader/training work plus O(P²) pair state/work per streamed block; retained training and optional dev metas prevent constant-space claim. LIS O(nlog n), UPGMA O(n³) distance visits, BTreeMap accounting O(tokens log vocabulary), rendering/output byte costs. Compact12C+17E+8T is conceptual bytes, not RAM.
composition
Real joins: bible::read_pool/sources text→corpus PVerse; main run_sys→measure::Run→Agg/VRec→block_measure and report::write; many→corpus/measure plus own pair metrics; pages receives wlc_jps VRec stage arrays from main and Hebrew encoder outputs. many::agree skips punctuation while block_meta includes it: their counts are not interchangeable. Report groups collapse layers. No invented call from report to many.
domain
Synthetic in-memory keys, residuals, metric records and permutations for executed evidence. Full CLI and real corpora/models/annotators/training/page/report effects are source-inspected only.
limitations
Finite admitted numeric domains and valid trees; each task owns its files; no claim about empirical corpus quality, persistent learning, or full pipeline correctness.
checks
- 15 exact native names twice debug/twice release, outcomes compared
- Five compiled behavior controls rejected and exact-restored with positive reruns
- Nine syntax projections generated twice and equal
- Local augmented registry and staged/range guards
catalog
behavior-complement
Paired aggregates exclude absent original source keys; diagnostic target reconstruction cannot imply normal exact translation. Page Meta cells excluded from page cost, while translation residual cells all count. Partial conditional exclusion over valid declared states.
reverse-reconstruction
Residual+decoder state reconstructs target layers conditionally; summary ratios admit many source texts, trees, gaps and sampling histories. Recovering source or meaning uniquely is unsupported. Preserve population masks, keys and residuals when diagnosing.
population-transport
All target VRec keys versus source-present Agg keys differ; missing whole blocks in many contribute absent/level accounting separately from cross population.
cost-denominators
Pair-local1000(reader+tail)/max(target_words,1), then arithmetic mean over pairs differs from pooled numerator/denominator; class sums and page filtered sums use distinct cell universes.
effect-order
Eval rejects train before output creation; both commands load model after output creation; report writes sequential artifacts, optional pages later. Source-derived effect order only, no rollback promise.
coordinate-join
VKey split depends on Kind; corpus joined heads add offsets preserving0; measure block sentinel contributes an additional offset. Unknown boundaries prevent unique reverse segmentation.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/translation-evaluation-pipeline.groupant/