Elamur Groupantsfish-pages-composition
The groupant beside tools/fish-pages/composition.groupant 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 derived | forward |
| ast | not materialized | |
| composition | partial source derived | composition |
| cost | partial source derived | cost |
| descriptant | partial source derived | domain |
| invariant | partial source derived | invariant |
| tests | finite checks | 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 |
|---|---|---|---|
behavior-complement | behavior_complement | source derived finite checks | behavior-complement |
reverse-reconstruction | reverse_reconstruction | source derived finite checks | reverse-reconstruction |
coverage-join | source_cardinality | source derived finite checks | coverage-join |
closure-join | process_effect_boundary | source derived finite checks | closure-join |
Model text
domain
Owned children are the pager library and CLI. External sibling dependencies are prepared-swarm history/execution and the configured native annotation executor; corpus/phase selection belongs to a separately pinned producer. Observations distinguish full intended source scope from ready pages and actual structural reviews.
forward
Join immutable global source intents to bounded producer admissions, actual Task.target_ids/gaps, one existing child step and durable page evidence. Source-ID counters compose across disjoint intents rather than multiplying batch counts. The cursor advances only after a source-bound seal and verified closed page history.
behavior-complement
A valid CLI invocation cannot substitute for an admitted page; a complete ready-page partition cannot substitute for a closed native attempt; a closed attempt cannot substitute for semantic truth. Missing pages wait and explicit preparation gaps remain pending review. Counterexample: exhaustion of an all-deferred plan is not full-corpus annotation. No child-local success proves the producer's held-out/source policy.
reverse-reconstruction
The aggregate tuple(validated,quarantined,deferred,rejected,cursor) has many compatible source/batch histories. Pinned intents, admissions and native receipt chains narrow those histories but do not reconstruct unique linguistic meanings or remote accounting. A checkpoint without valid seals cannot recover authority to skip work. Evidence: source inspection and finite source-count/restart/gap controls.
coverage-join
Type: source_cardinality. Global disjoint ownership plus each page partition implies every planned ID is accounted for as queued or an explicit gap. Native complete-target validation then contributes exact target-set cardinality to one review state. This conditional composition does not prove that the plan contains every desired corpus ID; producer completeness remains separately evidenced. Counterexample: a tiny plan can be internally complete while omitting most of a corpus. Evidence: duplicate/missing/foreign-source negative controls and distinct source/job counts.
closure-join
Type: process_effect_boundary. Safe parent snapshots require the imported raw task/profile and native process-closure predicates, current page audit, completed seal prefix and matching kind/config/plan. Fresh PAUSE differs from terminal proof requiring unheld leases. Counterexample: model process exit without an absent process group, or an orphan retained after STOP, does not permit quiescence. Finite fake Rust controls cover selected joins; no global exactly-once theorem.
invariant
All emitted review state remains training_authorized=false. Quarantine/defer/reject do not disappear into validated totals. The CLI and library preserve uncertainty instead of restarting paid work blindly. Parent source bindings own only the two local Rust children.
cost
Startup indexes global ownership; ordinary work scans only a bounded current page. Completed prefix audits occur on restart/read-only observation. Native execution latency and producer preparation cost remain separate and must be measured in deployment.
composition
CLI/control and library/source/evidence joins are conditional on conforming immutable producer artifacts and native execution. The mathematical model is partial and does not infer annotation accuracy from checked counts.
checks
Eight local fake Rust paging fixtures and thirteen existing prepared-swarm controls passed offline/locked j2; source/manifest binding and deterministic syntax guards are separate structural checks. No live-model, GPU, DB, training or full-corpus assessment evidence is asserted.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/fish-pages/composition.groupant/