cycle-pipeline

The groupant beside tools/cycle/src/pipeline.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	`model.md`
ast	partial	`syntax.json`
composition	partial	Composition
cost	partial	invariant-and-cost
descriptant	present	descriptant
invariant	partial	invariant-and-cost
tests	bounded or representation	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
`flight_reachability`	flight_reachability	partial source derived model	flight_reachability
`heldout_scope`	heldout_scope	partial source derived model	heldout_scope
`legacy_budget`	legacy_budget	partial source derived model	legacy_budget
`durable_attempt`	durable_attempt	partial source derived model	durable_attempt
`training_status_partition`	training_effect_partition	partial source derived compile checked model	training_status_partition
`training_status_preimage`	training_status_preimage	partial source derived compile checked model	training_status_preimage

Parts

Model text

descriptant

cycle-pipeline

Epoch 2026-10-31

Prospective inspection identified this relation before additions: Collect ready flights before feeding one next round, retain heldout/inflight exclusion set and artifact restart markers. Feed prepares draft/open parsers/context/prag then submits. Collect updates DB/prag, appends corpus, checks trees, optional enrichment/train, writes row. Legacy cascade chooses changed half then seeded hash remainder. Enrichment lock and attempt ledger reserve budget before teacher review/refine.

The production implementation is unchanged; only selected inline tests are appended. Source-bound syntax is generated twice and compared. Native checks exercise bounded helpers; entry orchestration, teacher calls, models and corpora remain source-derived. Partial syntax is not semantic proof. Preserve failure prefixes and ambiguity when refining this model.

Epoch 2026-11-01

Bound fresh input to the persistent reviewed train partition so the GP_ALWAYS path cannot bypass policy. No live teacher/trainer/GPU execution is part of this source change's tests.

Truthful training report — 2026-11-01

An audit-only collection could previously print GPU SLM fine-tuned even with training disabled. The forward effect model and training-status preimage distinguish four branches and expose the false implication from successful collection to training. The source now selects a report suffix from the actual branch, placing completed-here after successful train-command return and publication rename. It retains the original guard order, subprocess arguments, metric parsing and cycle.tsv row. Existing-candidate wording deliberately claims presence rather than verified integrity or new training; no GPU or quality inference is made.

The reporting change is compile checked and source reviewed without running collect or its effects. Prior helper tests retain their dated scopes. Old frozen-runtime logs are not rewritten, and this source change does not replace any active binary.

approximant

Forward relation

Collect ready flights before feeding one next round, retain heldout/inflight exclusion set and artifact restart markers. Feed prepares draft/open parsers/context/prag then submits. Collect updates DB/prag, appends corpus, checks trees, optional enrichment/train, writes row. Legacy cascade chooses changed half then seeded hash remainder. Enrichment lock and attempt ledger reserve budget before teacher review/refine.

Composition

Real calls join main parser/model/report helpers and db persistence. Model publication only after sh success. Legacy tier selection and enrich are conditional branches; independent lesson/twin tools are not pipeline callers.

Reviewed training — 2026-11-01

Trainer fresh selection uses db::blocks with sent_id boundaries and reviewed::training_allow, matching training_file's explicit train/dev split and heldout exclusion. Strict training rejects an unreviewed secondary DB before invoking training. Fine-tune base arguments still require caller provenance checks.

Training invocation observation — 2026-11-01

Successful collect reports the branch actually taken: disabled, no configured gparse model, an existing round candidate, or a completed training-command/publication prefix. Enabled training alone does not establish execution. The completion status is selected only after sh returns success and the candidate-directory rename succeeds; later row/report failures still prevent a final successful round report. Existing candidate status records the existing model.safetensors gate, not artifact validity or new training. Guard evaluation order and all training commands, numeric apply metrics and cycle.tsv fields remain unchanged.

catalog

Typed observations: cycle-pipeline

Subject: tools/cycle/src/pipeline.rs, SHA-256 f631f3949f282bea361d58e43e092b1c323b1ec697ce7ad871c3dfcbbf05f336. Domain: bounded source inputs, declared filesystem/environment state and normal allocations. Each view is a partial source-derived relation, not a universal proof. Observation is one helper call or the explicitly described trace. Checks are in tests.md; unexecuted effects remain gaps.

behavior_complement

No new feed while flight.len≥max. Successful reservation at used≥budget is excluded. However earlier DB effects are not rolled back by later collection failure; exactly-once database history cannot be claimed.