Elamur Groupantsmova-nautilus-sources
The groupant beside tools/nautilus/src/sources.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
model
mova-nautilus-sources
Descriptant
The sources as annotated units: command lessons (every explain/fact/question/option, id commands:<lesson>:<field>:<n>), the command corpus JSONL prompts, both annotated by en.bin; the Wikipedia corpus's ast.conllu sentences with sent_id and the article class. Source: tools/nautilus/src/sources.rs.
Forward approximant
lesson text -> (id, doc, text) per field and option (stars stripped); corpus JSONL -> (id, doc, text) per line with text; Annotator tokenize+annotate -> Unit words (form, lemma, UPOS); CoNLL-U -> Unit per sentence (multiword ranges and empty nodes skipped); files in name order.
Invariant
Directory order is sorted; ids are unique per source; nothing is written.
Cost
O(bytes) reading plus one annotation per commands unit.
Composition
Parent: nautilus. Parent: nautilus. State: Reads files only.
behavior_complement
Does not read lesson headers as text, deduplicate prompts, repair trees or read articles without ast.conllu. Evidence status: source-derived model with finite checks, not a proof.
reverse_reconstruction
Unit ids recover the lesson, field, question and option, or the corpus file and line, or the Wikipedia sentence. No unique inverse is claimed.
Epoch 2026-11-04
First model, written with the source; no prospective prediction is claimed beyond the checks.
Epoch 2026-11-04 (pool)
courses_raw reads school course folders: only explain, fact and exercise lines (LEARN_FIELDS); a course folder with a held-out name (test, exam, control, heldout, held-out) and a lesson file named held-out/heldout are never opened; lessons whose header declares CC BY-SA are reported. Red controls: test/exam lines read; held-out folder read.
Epoch 2026-11-04 (pool)
courses_raw reads school course folders: only explain, fact and exercise lines (LEARN_FIELDS); a course folder with a held-out name (test, exam, control, heldout, held-out) and a lesson file named held-out/heldout are never opened; lessons whose header declares CC BY-SA are reported. Red controls: test/exam lines read; held-out folder read.
Epoch 2026-11-04 (tier 1)
wikipedia also reads the large corpus's tier 1 when <corpus>/tier1/index.tsv exists: index rows in title order (title, shard, offset, length, …), each record one zstd frame at that offset, its conllu and class read as before. Doc ids are the title with _ and %2F (the folders' slug). Red control: a short index row fails (tier1_records_are_read).
Epoch 2026-11-04 (no assessment)
The command course reads only what is taught, like the course source: lesson_units uses LEARN_FIELDS (explain, fact, exercise and an exercise's options); test and exam lines and their options are never units, and a lesson file named held-out or heldout is not opened. The five-field list that read assessment is gone. Measured 2026-11-04: the commands draft (top 1000) held 985 examples from test/exam lines before and 0 after; 145 of 1000 words changed. Red controls: assessment fields read again; held-out lesson file opened (each red, restored).
tests
Checks
Command: 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. These are finite fixtures; no teacher or model was called; not a proof.
Pool epoch (2026-11-04): cargo test --release with the 0.4 en.bin, 26 unit + 2 English tests, two identical runs; red controls in the pool, improve and sources tests (each red, restored). Finite fixtures, not a proof.
2026-11-04 (no assessment): cargo test --release --locked (0.4 en.bin), tools/nautilus 38 unit + 2 English. sources::tests: lesson_lines_become_units (taught lines only; zorblax/quuxite of test and exam never units), the_command_course_reads_only_what_is_taught (test and exam lines and a held-out lesson file never read). Red controls: assessment fields read (two tests red), held-out file filter removed (red); each restored.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/nautilus/src/sources.rs.groupant/