Elamur Groupantsmova-hard-cases-judge
The groupant beside tools/hard-cases/src/judge.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 | Forward approximant |
| ast | partial | syntax.json |
| composition | partial | Composition |
| cost | partial | Cost |
| descriptant | present | Descriptant |
| invariant | partial | Invariant |
| tests | finite evidence | 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 |
Model text
Descriptant
Recheck of the list with the deployed gparse (gparse view: learned / reopened), the fixed measurement sample, measurement of a model (pass rate, drafts' rate, silver UAS/LAS on harvested trees, EWT dev LAS) and the watch loop. Source: tools/hard-cases/src/judge.rs.
Forward approximant
judge(tree, stanza) = roots | head_outside | grammar (world::induce::violations>0) | nonprojective | witness (UAS<0.8). sample: once, cases with drafts, measure split first then key, Stanza heads stored. measure(model) = gparse annotate (CPU) on the sample -> pass = no reason; attach() vs harvested trees; gparse eval on EWT dev. recheck = hard_case_list::recheck(gparse, all) with the judge; once per model id. watch: when the deployed weights' id differs from recheck.json -> measure then recheck.
Invariant
The sample file is never rewritten; a measurement of the same model and sample is byte-identical per case (CPU); recheck never touches the en view; cases without tokens stay unjudged.
Cost
One gparse CPU pass over the sample/list (~9 s for 131 sentences with EWT dev eval); one Stanza CPU pass per recheck.
Composition
Uses hard-case-list, policy, harvest annotations, gparse and the Stanza bridge. Parent hard-cases.
behavior_complement
Does not establish correctness (agreement with Stanza can be shared error); does not touch held-out test sets; does not change models. Evidence status: source-derived model with finite checks, not a proof.
reverse_reconstruction
Per-case verdicts are kept in measure/runs/<label>-<model>.json; the model is identified by its weights' hash, not recoverable from the verdicts.
Tests
judge::tests (2026-11-04): green tree; red: two roots, head outside, two subjects, witness disagreement; attachment counts. Live: same model measured twice identical per case; drafts 0/131 (red by construction); an older model 68/131 vs deployed 66/131 shows the instrument moves. Finite fixtures and live samples; not a proof.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/hard-cases/src/judge.rs.groupant/