📐nautilus

The groupant beside models/nautilus.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

Components

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.

Model text

model

nautilus

Descriptant

Nautilus (mova-nautilus): the dictionary maker, one of Mova's animals. Native dictionary layers from courses and corpora, English through the native deputy, a bounded standing loop with review-only proposals, and its teacher through its own mova-terminal behind a narrow interface (no model call in the tests).

Forward

sources -> annotated units -> layer::build -> render -> layer file; deputy prompt -> Hands(draft|stats) -> English; work: Job -> native or teacher::submit through a PrivateTerminal -> validate -> proposal -> journal.

Invariant

Native work is deterministic; every teacher question needs the operator's word, is journaled before and after, and is never asked twice; nothing is admitted by the animal.

Cost

Drafts O(tokens); one question per loop step at most.

Composition

Leaves are file groupants under tools/nautilus/src/*.rs.groupant and tools/nautilus/tests/*.rs.groupant. Correct leaves do not prove the wiring: the interaction checks are the English tests, which run the deputy, the handler and the real work together.

behavior_complement

No model calls in its tests; no admission; no slots; drafts only for configured layers; the Wikipedia layer stays outside the repository.

reverse_reconstruction

Layer files, the deputy ledger, the teacher journal and work.jsonl recover what was made, asked, answered, proposed and rejected; not unchosen units or question texts. Evidence status: source-derived with finite checks.

Epoch 2026-11-04 (pool)

Two children joined: mova-nautilus-improve (the layer source: weak rows of an existing layer asked in tens, proposals in the builder's formats, the layer file read only) and mova-nautilus-pool (a Crab pool member: the lease read between steps, a job's shared ledger, report and backlog files). Interaction: animal::Config::job builds a LayerJob or a DictJob for part of parts; work::Loop with a ledger makes a job movable between members without repeated questions; the counterpart is crab-pool, which writes the leases and reads the reports and backlogs. Licence classes are enforced by the configuration (shipped layers never take CC BY-SA, NC or research-only inputs) and by the data (a CC BY-SA header in a lesson or layer file). Finite checks only; red controls listed in the children's tests.

Epoch 2026-11-04 (senses)

Two children joined: mova-nautilus-senses (lemma matching of words in examples, junk and boilerplate, the pupils' numbered senses, native picks by lemma overlap, restatements, absent answers, the shared sense store) and mova-nautilus-replay (the loop measured on copies with earlier real answers). Interaction: DictJob and LayerJob plan every word through senses::plan; the loop writes Mova's picks next to the teacher's proposal and both feed the store (Job::learn), which every later layer and instance reads (a fixed sense is asked only as same); the configuration carries the store, the batch and the pacing. Measured 2026-11-04 by replay on a snapshot of 129 questions: 47 questions instead of 129, 23.4 words per question instead of 9.9, 578 question bytes per word instead of 771, 126 words settled natively, 107 of 1 030 teacher senses restated a numbered sense. Prospective: on live runs about one word in ten settled natively and about one teacher sense in ten marked as a restatement; to be corrected from the first live day.

tests

Checks

cargo test --manifest-path tools/nautilus/Cargo.toml --offline --locked with MOVA_OCTOPUS_TEST_MODEL=$MOVA_DATA/download/0/0.4/en.bin (15 unit + 2 English tests; two identical runs on 2026-11-04). Negative controls 2026-11-04 (each red, restored): no operator-word check; no terminal-id check; a limit not stopping the loop; stats drafting instead of reading. No teacher or model call; the private terminal is a fake in the tests.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/nautilus.groupant/