📐gparse-eisner

The groupant beside gparse/src/eisner.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

• gparse-eisner-dynamic-program
• gparse-eisner-backtrace

Model text

descriptant

Source-bound description

For square sc[h][d] with finite f32 intermediates, n=rows-1, complete/incomplete word spans combine smaller spans by f32 max-plus; root r maximizes sc[0][r]+C_right(1,r)+C_left(r,n). Explicit stack recovers n heads. len<=1 returns [], n=1 returns [0] before scores are read.

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.

approximant

Forward approximant

For square sc[h][d] with finite f32 intermediates, n=rows-1, complete/incomplete word spans combine smaller spans by f32 max-plus; root r maximizes sc[0][r]+C_right(1,r)+C_left(r,n). Explicit stack recovers n heads. len<=1 returns [], n=1 returns [0] before scores are read.

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

On admitted square scores with finite intermediate arithmetic, one root child, valid head coordinates, no self-head and projective tree. Real max-sum optimality is the mathematical recurrence model; f32 rounding/association is not certified equivalent to a real-valued global optimum. Strict > retains first ascending split/root under computed equal values. No NaN/infinite/ragged matrix guarantee.

Cost

O(n³) arithmetic and O(n²) score/backpointer storage plus O(n) output/stack; unchecked (n+1)² allocation arithmetic.

Allocation, finite arithmetic, platform dimensions and external failures are not universally certified.

composition

Composition

For square sc[h][d] with finite f32 intermediates, n=rows-1, complete/incomplete word spans combine smaller spans by f32 max-plus; root r maximizes sc[0][r]+C_right(1,r)+C_left(r,n). Explicit stack recovers n heads. len<=1 returns [], n=1 returns [0] before scores are read.

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

On admitted square scores with finite intermediate arithmetic, one root child, valid head coordinates, no self-head and projective tree. Real max-sum optimality is the mathematical recurrence model; f32 rounding/association is not certified equivalent to a real-valued global optimum. Strict > retains first ascending split/root under computed equal values. No NaN/infinite/ragged matrix guarantee.

For exact arithmetic (or an explicit order-preservation assumption for every candidate comparison), tree h defines score matrices whose tree sum dominates every projective single-root competitor, with implementation tie ordering on equality. For actual f32, the compatible set is defined by the rounded recurrence/backpointer decisions, not guaranteed identical to those real-score cones. Common per-dependent shifts preserve exact-real argmax; f32 equivalence additionally needs no rounding-order change. Singleton accepts every two-row matrix.

models

Typed model catalog

behavior_complement

Admitted finite-intermediate-domain output cannot contain two roots or an arc outside 0..n; scores into dependent zero do not affect output. Singleton ignores even missing inner rows; no score validation occurs on that branch.

reverse_reconstruction

For exact arithmetic (or an explicit order-preservation assumption for every candidate comparison), tree h defines score matrices whose tree sum dominates every projective single-root competitor, with implementation tie ordering on equality. For actual f32, the compatible set is defined by the rounded recurrence/backpointer decisions, not guaranteed identical to those real-score cones. Common per-dependent shifts preserve exact-real argmax; f32 equivalence additionally needs no rounding-order change. Singleton accepts every two-row matrix.

max_plus

Incomplete span I_left/right(s,t)=max_r(C_left(s,r)+C_right(r+1,t))+edge; complete spans close through an incomplete child. These are exact-real model equations; actual operations round in f32, and exact-real optimality needs preserved candidate ordering.

tie_policy

All-zero 3x3 gives heads [0,1]. First-ascending tie statements require every candidate sum encountered to be finite so initial -infinity is strictly improved; all-infinite candidates retain defaults instead.

domain_boundary

Zero rows/root-only and singleton bypass recurrence. Ragged matrices n>=2 may panic; nonfinite values can leave defaults and are excluded.

coordinate_transport

Inputs are head-major sc[h][d], unlike network arc tensor [batch,dependent,head]. Caller transposition is required.

tests

Checks and limits

eisner::tests::best_tree_with_one_root; eisner::wave30::empty_singleton_and_first_root_tie; decoder_accepts_no_words

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 — 2026-10-31

Source: gparse/src/eisner.rs

• eisner::tests::best_tree_with_one_root: twice debug, twice release, all four exits 0.
• eisner::wave30::empty_singleton_and_first_root_tie: twice debug, twice release, all four exits 0.
• decoder_accepts_no_words: twice debug, twice release, all four exits 0.
• synthetic_cpu_prediction_preserves_slots_and_fallback: twice debug, twice release, all four exits 0.

Controls

1 source-local compiled assertion mutants; 2 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

Finite tiny fixtures only; no exhaustive optimality oracle. F32 intermediate overflow and real-order preservation uncertified.

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.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/gparse/src/eisner.rs.groupant/