Elamur Groupantsqagen-fix-batching
The groupant beside qagen/src/fix.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 source model | forward |
| ast | partial | ../syntax.json |
| composition | partial source model | composition |
| cost | partial source model | cost |
| descriptant | partial source model | epoch-memory |
| invariant | partial source model | invariant |
| tests | partial source model | checks |
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 |
|---|---|---|---|
qagen-fix-batching-behavior-complement | behavior_complement | partial source model | behavior-complement |
qagen-fix-batching-reverse-reconstruction | reverse_reconstruction | partial source model | reverse-reconstruction |
qagen-fix-batching-preflight-order | preflight_order | partial source model | preflight-order |
qagen-fix-batching-legacy-domain | legacy_domain | partial source model | legacy-domain |
qagen-fix-batching-callback-count | callback_count | partial source model | callback-count |
qagen-fix-batching-batch-order | batch_order | partial source model | batch-order |
Model text
forward
try_new checks alignment then delegates stable grouping; prompt numbers each paragraph's items; Job adapter uses the same vectors.
invariant
For checked input every item appears in exactly one matching paragraph batch and has in-range anchors; original per-paragraph item order survives.
behavior-complement
try_new never drops orphan items or duplicates their emission via duplicate paragraphs. Legacy new remains unchecked; missing direct Job batch access can panic.
reverse-reconstruction
Grouping loses input interleaving between paragraphs; multiple item permutations across distinct paragraphs yield the same prompt. Prompt text does not recover reason metadata.
cost
Validation O((P+I) log(P+I)+anchors); grouping expected O(P+I) key hashing plus BTreeMap factors. Prompt item filtering can be O(P*I), not optimized here.
composition
Validation closes missing/duplicate/batch/anchor preconditions before inherited grouping. Prompt and gate agree on stable item numbering when using checked batches.
preflight-order
Duplicate paragraph first; per item missing paragraph, duplicate item, mismatched batch, invalid anchors. Empty paragraph/item collections accepted.
legacy-domain
new retains previous HashMap behavior; callers needing validation choose try_new.
callback-count
Job check calls gate once; it does not invoke teacher or lemma callbacks.
batch-order
BTreeMap orders batch IDs; input paragraph order is retained within each batch.
checks
Synthetic embedded Rust tests, 1024 independent coverage cases, preflight sentinel artifacts, callback trace and compiled branch/arithmetic controls; debug and release twice, model-loading test explicitly skipped. See parent evidence.md for observed outcomes, not proof.
epoch-memory
Prospective model was saved before repair on 2026-10-30. Old source failed four library witnesses and one CLI witness. The tie control initially survived because [2,1,2] selects 2 for both first and last ties; the corrected [2,1] witness distinguishes them. Preserve this observation: duplicate-endpoint fixtures can hide tie regressions.
references
Original project implementation and rule v2: docs/context.md, qagen/src/gate.rs and qagen/src/run.rs. Model conventions: docs/groupants.md, docs/mathematical-modeling.md and docs/mathematical-modeling.md. No external solution code copied.
limitations
Finite local tests and partial syntax projections, no semantic AST or universal proof. Allocation failure and callback panic are outside successful-return model; filesystem races and write failures can interrupt publication. No live model/teacher/database/corpus used. No useful certified numeric surrogate is attached to these categorical/effectful interfaces; no invented error budget, substitution or speed claim.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/qagen/src/fix.rs.groupant/batching.groupant/