Elamur Groupantsen-tag
The groupant beside en/src/tag.rs 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 | approximant |
| ast | partial | syntax.json |
| composition | partial | composition |
| descriptant | present | descriptant |
| invariant | partial | invariant |
| tests | executed bounded checks | tests |
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 |
interpolation_boundary | interpolation_boundary | partial source derived model | interpolation_boundary |
suffix_backoff | suffix_backoff | partial source derived model | suffix_backoff |
emission_precedence | emission_precedence | partial source derived model | emission_precedence |
viterbi_ties | viterbi_ties | partial source derived model | viterbi_ties |
Parts
Model text
descriptant
en-tag
TnT trigram tables, lexical emissions, suffix model, Viterbi
Epoch 2026-10-31: read complete source before prospective model and bounded tests. Production code retained byte-for-byte; only inline synthetic checks added. Coverage is partial, not a proof.
2026-11-04 epoch (en-morph): the unknown-word dictionary hint allows the derivation tags (derive::hint) of a lower-case form the dictionary does not know. Source e2175f4ca53b; earlier receipts and evidence refer to the previous source hash and are kept as history.
2026-11-04 epoch (lexicon growth): the unknown-word dictionary hint takes lexext::hint(word, capitalised) first, then the derivation's tags for a lower-case word.
approximant
Forward relation
Tagged sentences only; double BOS then final BOS. Interpolation ties trigram then bigram; rare<=10 suffix max10. Exact then lowercase known lexicon bypasses hints. Unknown suffix stops at first absent; dictionary soft/hard then seen positive emission candidates.
No executable Rust surrogate attached: no demonstrated metric for preserving these categorical, stateful or effect observations. Supported future caps are 1/2/5/10%, not a claim of current error.
2026-11-04: unknown(w): allowed tags = dictionary analyses ∪ (no dictionary form ∧ not capitalised ? derive::hint : ∅).
invariant
Domain, invariant and cost
Tagged sentences only; double BOS then final BOS. Interpolation ties trigram then bigram; rare<=10 suffix max10. Exact then lowercase known lexicon bypasses hints. Unknown suffix stops at first absent; dictionary soft/hard then seen positive emission candidates. Empty tagging empty; empty training has NaNs/no support and nonempty inference may panic. Deserialize and Soft multiplier have no validation. Train O(T+NT^3+suffix work); decode O(nK^3) time/O(nK^2), NT fixed; unchecked counters. Inputs must satisfy the indexing/generated-data assumptions stated by the typed catalog; allocation failure and overflow beyond bounded fixtures remain untested.
composition
Mathematical combiner
Source bytes and twice-generated syntax constrain this model. Forward relation: Tagged sentences only; double BOS then final BOS. Interpolation ties trigram then bigram; rare<=10 suffix max10. Exact then lowercase known lexicon bypasses hints. Unknown suffix stops at first absent; dictionary soft/hard then seen positive emission candidates. Reverse relation: A sequence has all words/parameters for which its score maximizes under strict-greater first-candidate ties; aggregated training counts lose original sentence ordering. The ordered typed family records independent exclusions, compatible antecedents and source-specific boundaries. Fragment ranges compose through actual control flow; they do not cover all helpers. Epoch memory: empty learned support, lossy projections and non-atomic effects cannot inherit stronger downstream guarantees.
models
behavior_complement
Empty tagging empty; empty training has NaNs/no support and nonempty inference may panic. Deserialize and Soft multiplier have no validation.
reverse_reconstruction
A sequence has all words/parameters for which its score maximizes under strict-greater first-candidate ties; aggregated training counts lose original sentence ordering.
interpolation_boundary
Deleted interpolation chooses q3 on ties then q2; no counts makes lambda division0/0 and NaN transition cells.
suffix_backoff
Suffix smoothing (q+theta*p)/(1+theta) stops at first missing suffix, cap flag distinguishes uppercase first scalar.
emission_precedence
Known exact/lowercase lexical entries return before dictionary; hints only modify unknown probabilities; hard cannot add unseen classes.
viterbi_ties
Viterbi sums transition+emission and final BOS; strict > retains first finite maximizing predecessor; candidate emptiness is not validated.
tests
Actual source checks
Named check: tag::wave27_tests::tiny_known_emissions_and_empty_boundary. Command: cargo test -p en --lib wave27_tests -- --test-threads=1 (also --release). Four runs: each 10 passed, 0 failed, numeric exit0; sorted named outcomes equal. This file contributes one named check per run. Control present for this source: false. Hash zeroed byte contribution, text changed empty chrF to1, lexgen removed three-cell PastPart copy: each compiled and failed one assertion (exit101), then exact source restoration and ten-test positive rerun (exit0). Other sources have no compiled mutation claim. See evidence.json and external octopus-wave27-lexical-2026-10-31 receipts. Static built dictionary read is part of the linked library; no corpus or external services used. Unknown support, malformed model/blob, resource failures, cycles and exhaustive Unicode behavior are unexecuted gaps.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/en/src/tag.rs.groupant/