📐parsecmp-main

The groupant beside tools/parsecmp/src/main.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

Components

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.

Parts

• parsecmp-main-bracket-intersection

Model text

descriptant

Descriptant

read_conllu ignores read errors, short/range/dotted rows, defaults malformed head to u32::MAX and lowercases relation. Score counts head matches, UD-only base LAS and gold-PUNCT exclusion. spans strips functional label suffixes, excludes root/preterminals and sorts half-open word spans. bracket_f1 counts multiset intersection. Main aligns by counts/order, requires two UD voters, ties smaller head, flags changed head at >=0.75 support; writes three reports sequentially. agree uses first system punctuation mask for all pairs.

Epoch 2026-10-31

Prospective source reading identified conditional input alignment, information loss and failure order before tests. Existing production code is retained; added tests use actual helpers. Finite results update this source-bound model, not all parser/corpus semantics.

No executable percentage surrogate has a justified observation metric for these categorical/effectful APIs. Future profiles require metric, domain and real-reference evidence at caps 1/2/5/10%; no higher cap. Syntax is partial lexical Rust projection without macro expansion or semantic resolution.

approximant

Forward approximant

read_conllu ignores read errors, short/range/dotted rows, defaults malformed head to u32::MAX and lowercases relation. Score counts head matches, UD-only base LAS and gold-PUNCT exclusion. spans strips functional label suffixes, excludes root/preterminals and sorts half-open word spans. bracket_f1 counts multiset intersection. Main aligns by counts/order, requires two UD voters, ties smaller head, flags changed head at >=0.75 support; writes three reports sequentially. agree uses first system punctuation mask for all pairs.

Composition

Child fragments model one narrow helper each; their correctness alone does not establish orchestration.

The containing source wires this scoped relation to its callers. Treeagree proposal→gate→commit and treebanks parse→match→collect→render are actual joins; parsecmp is an independent diagnostic path. The external group model distinguishes these tags.

catalog

Typed models for parsecmp-main

Subject: source SHA-256 8d5c1b3d4b13c46e7f0b31fad87b1197b29574d987a2ea1c89ed012961a9a9e7; manifest supplies exact source/byte scope. Domain: admitted states in forward model, nonoverflow arithmetic and successful allocations. Observation: one terminating invocation, except explicitly described file trace. Source-derived partial relations; named finite tests constrain selected cases only. Each entry holds arguments and environment fixed except its stated intervention.

behavior_complement

Non-UD systems never contribute LAS or majority votes. Span duplicates contribute at most minimum multiplicity. Token count alignment does not validate matching forms/ids or safe heads.

Operator: inner exclusion on admitted domain. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

reverse_reconstruction

Scores constrain numbers of matching positions but cannot recover trees; spans lose words, POS/preterminal labels and root wrapper; multiset F1 loses identities of missed spans.

Operator: compatible antecedent relation, retaining ambiguity. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

punctuation_mask

Score excludes by gold UPOS only; agree applies first parser mask even to later pair comparisons, so swapping parser order can change denominator.

Operator: specialized observation relation under the stated fixed inputs. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

zero_denominator

pct(0,0) returns em dash; agree reports 0.00 with max(1) denominator; constituency empty pair F1 is zero.

Operator: specialized observation relation under the stated fixed inputs. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

multiset_join

hit=sum_s min(count_a(s),count_b(s)); returned denominators preserve each full multiplicity.

Operator: specialized observation relation under the stated fixed inputs. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

read_failure

Missing file becomes empty sentences, indistinguishable from empty file at read_conllu boundary; main can still fail on course read.

Operator: specialized observation relation under the stated fixed inputs. Evidence: source branch derivation plus tests.md; unexercised branches remain hypotheses. Oracle: evaluate the scoped actual helper on admissible fixtures and compare this observation. No universal proof.

tests

Bounded evidence — 2026-10-31

Source tools/parsecmp/src/main.rs at SHA-256 8d5c1b3d4b13c46e7f0b31fad87b1197b29574d987a2ea1c89ed012961a9a9e7. Selected actual-source tests in parsecmp:

• wave33_tests::score_and_bracket_multiset_boundaries
• wave33_tests::conllu_fixture_filters_and_missing_file

These names passed twice in debug and twice in release; the complete treebank/comparison slice has 25 target-qualified names and identical sorted outcomes in all four runs. Ten compiled behavior mutations (one per owned source) failed named runtime assertions; exact original/restored SHA-256 equality and positive reruns were checked. Zero assertion-only mutants, zero surviving exploratory mutants, zero production fixes. Fresh raw logs, numeric exits, commands and mutation hashes are external in the wave33 trees receipt set.

Scope and gaps

Only named finite assertions cover selected inputs; other branches and invalid-domain behavior remain source-derived. No corpus, network, downloads, annotation model, live teacher, full release build or site execution. The treebanks join is entirely in memory and parsecmp uses a tiny uniquely named temporary directory. Treeagree Registry::load uses the embedded UD registry. Tests are finite CPU observations, not universal semantic proof, title-match quality estimates or corpus accuracy.

invariant-and-cost

Invariant and cost

Read O(bytes); spans O(chars+spans log spans) memory O(chars+depth+spans), bracket intersection O((a+b)log distinct). Main pairwise comparisons O(system²*tokens), holds all outputs; no parser/model invocation. Numeric sums assume no overflow.

Non-UD systems never contribute LAS or majority votes. Span duplicates contribute at most minimum multiplicity. Token count alignment does not validate matching forms/ids or safe heads.

The complement and specialized catalog constrain preserved coordinates and effect boundaries. Empty/malformed domains are only guaranteed where explicitly admitted. No whole-program cost bound follows from a helper bound.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/parsecmp/src/main.rs.groupant/