Elamur Groupantsmova-hard-cases-harvest
The groupant beside tools/hard-cases/src/harvest.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
Validated Fish replies of the lane's waves -> learn/annotations.jsonl and per track/split CoNLL-U and pragmatics TSV, review-only silver. Source: tools/hard-cases/src/harvest.rs.
Forward approximant
for wave in order, pager root, job in sorted order: accept when outer and inner receipts are validated_review_only (review_only, not training-authorized), reply binds task id, each target tree has the target id and its context text -> first annotation per target (minus held-out texts) -> write(annotations, track/split files, manifest).
Invariant
Only ledger targets with status queued are kept; one annotation per target; files are rewritten whole and atomically; nothing is gold or training-authorized.
Cost
O(jobs + reply bytes).
Composition
Reads fish-pages/fish-swarm job layouts (pages/<intent>/run/jobs/<job>/); feeds witness, eel and judge. Parent hard-cases.
behavior_complement
Does not judge annotation correctness, merge two replies or repair trees; quarantined, uncertain or unbound jobs are skipped with their reason counted. Evidence status: source-derived model with finite checks, not a proof.
reverse_reconstruction
Each annotation keeps its wave, job, receipt hash, source id and key, so the reply and the case can be found again; the other replies for the same target are not kept in learn/.
Tests
harvest::tests (2026-11-04): a bound validated reply counts; red: a quarantined job and a tree whose text is not its target's; learn files split by track and split. Finite fixtures and live samples; not a proof.
This page in Markdown ā https://md.groupants.elamur.ai/free/en/mova/tools/hard-cases/src/harvest.rs.groupant/