Movamova-enrich — grounded corpus enrichment and bounded self-study
mova-enrich turns a cycle round's agreed annotations (pre/agreed.conllu, pre/prag-agreed.tsv, world.tsv) into reviewed teaching material: a draft dictionary supplement, a books lesson, exercises, tests, exams and course-gap candidates. Mova prepares bounded proposals and course excerpts; the teacher (Codex CLI through mova-teacher) reviews them; Rust gates validate and stage the result. It also runs Mova's own bounded study/propose/review/learn/assess step (improve).
How it fits the training: Training Mova. The self-study step: The school.
Build
cargo build --release -p mova-enrich
Commands
Every output path must be new; nothing is overwritten.
| Command | What it does |
|---|---|
prepare ROUND DICTIONARY SCHOOL NEW_RUN MAX_SENTENCES | build the source-bound packet and review prompt from a cycle round |
review RUN | one teacher call on the prepared prompt |
stage RUN | validate the review and stage the drafts |
run ROUND DICTIONARY SCHOOL NEW_RUN MAX_SENTENCES | prepare, review and stage in one go |
check-source RUN ROUND DICTIONARY SCHOOL | check that the dictionary, the school lessons and the agreed trees still match the packet's fingerprints |
status RUN / profile RUN | print the run's state / prompt and packet sizes as JSON |
require-staged RUN | exit non-zero unless the run is a current, validated staged result |
refine OLD_RUN NEW_RUN | one targeted repair of a review whose only failure is an over-long answer span |
restage OLD_RUN NEW_RUN | stage an older run again under the current validators |
simulate NEW_JSON | write the artifact-state model check (completion needs all three artifact gates) and an idealized worker/wave grid; not a runtime forecast |
improve STAGED NEW_ROOT [BASE_MODEL] | the bounded study → propose → review → learn → assess step |
study STAGED NEW_STUDY [BASE_MODEL] | the study part of improve alone |
prepare-next STAGED STUDY NEW_REVIEW | prepare the next review from a study |
finish-improve ROOT | finish an interrupted improve from its saved parts |
study-exam STATE EXAM NEW_REPORT | answer an exam file from a saved retained state (fresh process, no lesson replay) |
Environment
| Variable | Default | Meaning |
|---|---|---|
ENRICH_MODEL | gpt-6.1-terra (via MOVA_TEACHER_MODEL first) | teacher model |
ENRICH_PROMPT_MAX_BYTES | 150000 | refuse a review prompt above this size |
ENRICH_CURRENT_DICTIONARY (or CYCLE_ENRICH_DICTIONARY) | — | the current dictionary whose meanings must be preserved |
Effort is medium. The binary prints wall_seconds on stderr.
Limits
- Lexical overlap retrieves related lessons; it does not prove a missing topic.
- Fiction does not become fact: narrative-driven additions and course gaps wait for independent evidence.
- A staged result is a draft for admission, not admitted knowledge; a whole pass is never marked seen after a partial failure.
- Different test scores or more frames alone are not evidence of improvement; compare the same questions before and after.
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/tools/enrich/