Elamur Groupantsgparse-data
The groupant beside gparse/src/data.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 |
| Cost | symbolic source derived | Invariant |
| Descriptant | present | Descriptant |
| Evidence | executed finite checks | Evidence |
| 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.
| relation | type | label | where |
|---|---|---|---|
| Behavior complement | behavior complement | partial source derived model | Behavior complement |
| Reverse reconstruction | reverse reconstruction | partial source derived model | Reverse reconstruction |
| Coordinate transport | coordinate transport | partial source derived model | Coordinate transport |
| Failure precedence | failure precedence | partial source derived model | Failure precedence |
| Tie policy | tie policy | partial source derived model | Tie policy |
| Stochastic state | stochastic state | partial source derived model | Stochastic state |
Parts
Model text
Descriptant
Source-bound description
read scans UTF-8 lines, ignores short (<8 columns), range and decimal-ID rows, retains first 8 fields, parses invalid head as zero and emits only nonempty sentences. write renumbers words and discards original DEPS/MISC/roles. Vocs lowercases word counts with lexical ordering; tags preserve first-seen order. batch_prag uses T=1+max(length), row=k*T+i, key root/real=1, mask real=1; gold heads clamp to sentence length. plan/plan_of sort (length,index), flush only a nonempty batch exceeding tokens. Lex::fill precedence is EModE table, form/tag frequency maximum with lexicographically smaller tie, then tag-frequency features and morphology. inference_cost sums BT, BT² and (n+1)³ for 0<n<max_len.
Approximant
Forward approximant
read scans UTF-8 lines, ignores short (<8 columns), range and decimal-ID rows, retains first 8 fields, parses invalid head as zero and emits only nonempty sentences. write renumbers words and discards original DEPS/MISC/roles. Vocs lowercases word counts with lexical ordering; tags preserve first-seen order. batch_prag uses T=1+max(length), row=k*T+i, key root/real=1, mask real=1; gold heads clamp to sentence length. plan/plan_of sort (length,index), flush only a nonempty batch exceeding tokens. Lex::fill precedence is EModE table, form/tag frequency maximum with lexicographically smaller tie, then tag-frequency features and morphology. inference_cost sums BT, BT² and (n+1)³ for 0<n<max_len.
Domain: valid source-specific dimensions and resources stated in invariant; no automatic production substitution. Numerical observations are finite checks, not an end-to-end relative bound. Executable Rust surrogate: unsupported gap; existing exact tests are not surrogates. Future metric-specific caps may be 1/2/5/10%, never above 10%.
Invariant
Invariant and cost
Root features [2,1,2,8]; role_arc[(k*T+d)*T+p]=r, first nonzero filler role retained and predicate flag adds 8 at most once for valid 1..=n roles. Empty shape is punctuation 6, empty Voc::load has one empty name; word hashing uses wrapping FNV and buckets 3..AFFIX. Valid roles are required for manually constructed Sent. No structural UD-tree validation.
Cost
O(input bytes + vocabulary log V + sort S log S + BT² + word characters) time, O(BT²+retained words/vocab) storage; lexical learning retains distinct form/tag/feature keys. Count/usize products and u32 narrowing are unchecked.
Allocation, finite arithmetic, platform dimensions and external failures are not universally certified.
Composition
Composition
read scans UTF-8 lines, ignores short (<8 columns), range and decimal-ID rows, retains first 8 fields, parses invalid head as zero and emits only nonempty sentences. write renumbers words and discards original DEPS/MISC/roles. Vocs lowercases word counts with lexical ordering; tags preserve first-seen order. batch_prag uses T=1+max(length), row=k*T+i, key root/real=1, mask real=1; gold heads clamp to sentence length. plan/plan_of sort (length,index), flush only a nonempty batch exceeding tokens. Lex::fill precedence is EModE table, form/tag frequency maximum with lexicographically smaller tie, then tag-frequency features and morphology. inference_cost sums BT, BT² and (n+1)³ for 0<n<max_len.
Scoped children
The byte-scoped children project distinct internal relations through the parent syntax file. The parent combines them in actual source order. Remaining declarations/tests outside child ranges belong to this parent. Their correctness does not establish external caller preconditions.
Interfaces and limits
Root features [2,1,2,8]; role_arc[(k*T+d)*T+p]=r, first nonzero filler role retained and predicate flag adds 8 at most once for valid 1..=n roles. Empty shape is punctuation 6, empty Voc::load has one empty name; word hashing uses wrapping FNV and buckets 3..AFFIX. Valid roles are required for manually constructed Sent. No structural UD-tree validation.
A parsed Sent has all UTF-8 antecedents differing in ignored rows, extra columns and invalid heads mapping to zero; write is a canonical projection, not original-text recovery. A batch has compatible original forms colliding in lowercase vocabulary/affix hashes and compatible dropped word IDs; fixed Vocs and no dropout do not make it injective.
Models
Typed model catalog
Behavior complement
No range/decimal row can become a word through read. Invalid numeric head cannot survive as an error: it becomes 0. Budget zero does not exclude nonempty singleton plans; absent sentence IDs do not clear old roles, while matching IDs replace them. Unknown lexical relation becomes its known base or dep unless relation set is empty.
Reverse reconstruction
A parsed Sent has all UTF-8 antecedents differing in ignored rows, extra columns and invalid heads mapping to zero; write is a canonical projection, not original-text recovery. A batch has compatible original forms colliding in lowercase vocabulary/affix hashes and compatible dropped word IDs; fixed Vocs and no dropout do not make it injective.
Coordinate transport
Local head h becomes flat k*T+h only at model loss/relation boundary; rows exclude root/pad but keys include root.
Failure precedence
read checks row width and ID markers before indexing; attach_roles parses row count/role/numbers before bounds filtering; invalid manual roles can panic during batching.
Tie policy
Word vocab counts iterate BTreeMap sorted lowercase forms; Lex::top chooses smallest equal-count key; role filler preserves first accepted role.
Stochastic state
Rng zero is absorbing; unit uses top 24 bits divided by 2^24. Drop branches draw only when drop>0, role vectors nonempty, and prag_drop>0; no guarantee for nonfinite probabilities.
Tests
Checks and limits
data::tests::features_line_up_with_words; data::cost_tests::cost_counts_real_batch_shapes_and_filters_skipped_sentences; data::wave30::projection_roles_and_zero_budget
Selected commands run twice debug and twice release; exact receipts and source-to-test mapping retained in wave30 gparse external run directory. Tests do not certify allocation failure, all NaN/overflow inputs, trained-model accuracy or corpus timing. Compiled mutants cover selected boundaries; per-source coverage is recorded in the external map, not presumed.
See source execution evidence for exact names and unexecuted paths.
Evidence
Execution evidence
Source: gparse/src/data.rs
- data::tests::features_line_up_with_words: twice debug, twice release, all four exits 0.
- data::cost_tests::cost_counts_real_batch_shapes_and_filters_skipped_sentences: twice debug, twice release, all four exits 0.
- data::wave30::projection_roles_and_zero_budget: twice debug, twice release, all four exits 0.
- training_attention_keeps_gradients_and_inference_control_breaks_them: twice debug, twice release, all four exits 0.
- synthetic_cpu_prediction_preserves_slots_and_fallback: twice debug, twice release, all four exits 0.
- synthetic_single_optimizer_step: twice debug, twice release, all four exits 0.
Controls
1 source-local compiled assertion mutants; 3 observed control tests including dependency mutations. Mutant compile exit0, assertion exit101, byte-exact restoration and positive rerun exit0. Per-source controls are listed in external source-test-map.json, not implied for every source.
Limits
Lex helpers unexecuted in selected checks; invalid manual role indices, large integer products and RNG distributions not exhaustively tested.
CUDA softmax comparisons use absolute tolerance2e-6. GELU grid maxima observed: values0.0004734993, derivatives0.00086838007, finite difference0.0000005534244635896357; respective strict caps5e-4/1e-3/1e-5. These are finite observations, not percentage profiles or throughput claims.
Receipts: external wave30 gparse run native-commands.json, results.json, controls.json and source-test-map.json. Sorted exact test names/outcomes match per profile; numeric diagnostics and timings need not be byte-identical. Strict scoped library Clippy reports ten pre-existing production lints; appended test blocks were formatted independently without production reformatting.
History
2026-10-31 · descriptant
Epoch 2026-10-31
Prospective model was recorded before edits. Added tests use bounded synthetic fixtures; production algorithms unchanged. Coverage is partial, not proof. No trained corpus, live teacher or throughput experiment was executed.
2026-10-31 · evidence
Heading ### Execution evidence — 2026-10-31 stays in evidence as ### Execution evidence.
2026-11-04 · descriptant
Epoch 2026-11-04
Word now keeps MISC (column 10, _ when absent); write still emits _ for DEPS and MISC, so annotate output is unchanged. vote parses Vote, VoteHead (defaults to Vote) and VoteAlt=h:s,… from MISC, clamps finite shares to [0,1] and returns None without a finite Vote. ngrams lowercases the word, adds boundary marks and hashes its 3- to 6-grams (shorter n first, at most NGRAM_CAP=32) by salted FNV-1a into 1..buckets. batch_full is the general batch: batch_prag and batch are its ngram=0 wrappers, batch_for its inference form; it adds ngrams (B·T·g, g = max per-word count, at least 1 when on), per-row arc_conf (VoteHead or 1), rel_conf (Vote or 1) and alts (other heads inside the sentence, not the gold head). Prediction of the earlier inputs: no change to inp, rows, gold or the RNG sequence for any ngram value or MISC content (checked by ngrams_and_votes, which compares inp with and without n-grams).
2026-11-04 · behavior_complement
Since 2026-11-04: a word without a finite Vote is never marked uncertain (arc_conf = rel_conf = 1); an alternative head outside 0..n or equal to the gold head never reaches alts; ngrams never yields bucket 0 for a real n-gram, and root/padding rows never carry n-grams.
2026-11-04 · reverse_reconstruction
Since 2026-11-04: n-gram buckets are a lossy many-to-one image of the lowercased spelling (hash collisions, the 32-gram cap, case folding); vote shares carry two decimals, so the (weight, version) histogram behind them is not recoverable.
2026-11-04 · tests
Since 2026-11-04: data::tests::ngrams_and_votes (n-gram count, case folding, shared n-gram bucket, cap, vote parsing, batch confidence rows, n-gram placement, unchanged inp).
Provenance
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/gparse/src/data.rs.groupant/