Elamur Groupantscommand-school-corpus
The groupant beside tools/octopus/src/commands.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 derived | Forward |
| ast | partial | syntax.json |
| composition | partial source derived | Composition |
| cost | partial source derived | Cost |
| descriptant | partial source derived | Descriptant |
| invariant | partial source derived | Invariant |
| tests | finite native 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 checked | behavior_complement |
reverse_reconstruction | reverse_reconstruction | source derived finite checked | reverse_reconstruction |
Model text
model
Command school corpus and exam — source-bound model
Descriptant
commands.rs is the start of the command school: an anonymized corpus of commands sent to Crab and an exam that
replays reviewed commands through Mova's understanding in plan mode (nothing executed). As in the primary school, a
command an arbitrary Mova-based animal does not understand signals a gap in the learning algorithm, not data to
memorize. Lessons and a command dictionary follow the primary-school pipeline later.
Forward
anonymize: whitespace tokens whose punctuation-trimmed core is an absolute or home path → <path>, contains :// →
<url>, looks like an e-mail → <email>, is a ≥12-char hex/dash run with ≥4 digits or a ≥5-digit number → <id>;
everything else (words, role names, small numbers) unchanged. record: append {text, priority, observed, origin}
(JSONL, 0600). exam: for each entry with expected, compare reading(review(text)) (plan capabilities in order, or
refusal stage) and count by kind; unreviewed entries are only counted.
Invariant
anonymize is idempotent; the exam never executes a capability (it calls review only); observed readings never become
expectations automatically — expected is written by review.
Cost
Linear in prompt length for anonymize; one native parse per examined entry.
Composition
The crab binary records every ask (optional corpus path on serve) and exposes crab exam and crab audit.
behavior_complement
Not anonymized: names of people or organisations written as plain words, short numbers, relative paths; the audit only checks idempotence of these rules, not semantic privacy. Not done: lesson generation, dictionary, training.
reverse_reconstruction
From an anonymized text the original paths, ids, e-mails and URLs are not recoverable (many-to-one); the reading recovers which configured capability or refusal stage the understanding produced, not the execution outcome.
Epoch 2026-11-04: Executed failures, review loop and generated exercises
Reading::ExecutedFailed{capabilities} marks an understood plan whose execution failed; it is never examined as a refusal and cannot be a reviewed expectation. review_queue lists observed texts without a reviewed expectation; promote admits reviewer decisions only when observed, privacy_review=passed, anonymized, a reading (plan/refusal) and not yet reviewed — never training. exercises derives determiner variants and same-shaped other-entity variants (same handler and parameters except the target; a headed name only within its configured group) from reviewed single-capability plans. group_candidates groups legacy and deduplicated review files read-only.
checks
Checks (2026-11-03)
anonymize keeps English and hides path/id/e-mail/URL, idempotent; audit refuses raw paths. Live: 21 reviewed seed commands examined twice with identical output, 21/21 after correcting one refusal-stage expectation.
2026-11-04: unit tests for determiner toggling, promote gate (privacy, observation, execution outcome, duplicates), review queue and exercises with a negative control; live: seed 52/53 (the one failure is a reviewed refusal now understood: "Is the documenter running?"), 123 generated exercises from the seed all pass.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/octopus/src/commands.rs.groupant/