Elamur Groupantsmova-hard-cases-witness
The groupant beside tools/hard-cases/src/witness.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
Stanza, Trankit and spaCy trf through the existing neural runner on a track's training trees under the shared GPU lease; # text attached only where id and every token form agree; agreed-input.conllu and parsers/out/*.conllu for the Eel's tree agreement. Source: tools/hard-cases/src/witness.rs.
Forward approximant
if witness.json binds sha256(train) -> return it; else write in/ids.txt, in/tokens.txt; with flock(gpu_lock) run neural.py for the three parsers; align each output by index (same id and forms) -> keep the records all three align; write agreed input, filtered parser outputs, witness.json.
Invariant
A record enters the agreed input only when all three parsers align on id and forms; parser output rows are unchanged except the inserted # text line; no witness is manufactured.
Cost
Three neural parser runs (GPU); measured 38 s for 3 sentences including model loading.
Composition
Feeds the Eel engine's agreement step (treeagree agree reads parsers/out). Parent hard-cases.
behavior_complement
Does not run without the GPU lease (waits up to lease_wait_seconds), does not retry a failed parser, does not repair tokenization. Evidence status: source-derived model with finite checks, not a proof.
reverse_reconstruction
The raw runner output stays in witness-run-<hash>/out; the filtered files plus the missing list recover which records each parser aligned.
Tests
witness::tests (2026-11-04): text attached only to aligned records; red: a different tokenization and a wrong id do not align. Live: three sentences, all aligned; treeagree agree accepted the output. Finite fixtures and live samples; not a proof.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/hard-cases/src/witness.rs.groupant/