AGENTS.md — Mova

These instructions are for AI agents (and people) working on Mova.

First read Gali and enter the mode. It is a working mode for an AI that builds with a human operator. Do not just skim it: step in and work.

What Mova is

Mova is a small, transparent language model written in Rust. English comes first. It parses into Universal Dependencies trees, reads the trees into frames and a world of the text, and answers from that world. Every step is deterministic and can say why. A large model is only a teacher: it never answers at test time. Start with Mova and Architecture.

Working rules

• English everywhere in code, comments, docs, commit messages and data. Talk to the operator in their language.
• Mechanics in Rust. Code chews the data, not your tokens. Avoid bash or Python improvisation; the few Python bridges, to the neural parsers and the pragmatic witnesses, are the exception.
• Compose from what exists. Grep the repo before you write new code.
• Deterministic. Use BTreeMap/BTreeSet wherever order decides anything. Before you claim a gain, run twice and check the two runs are identical.
• Measure before you claim. Keep held-out sets held out: the SVAMP test runs once per release, and decisions are made on the two-fold check (cv). State every number with its date.
• Negative control. A gate that cannot show red is not a gate.
• Fail-fast. No blind retries on expensive steps. New directions and expensive runs (large-model batches, long trainings) need the operator's explicit word.
• Two execution modes (Execution modes: CPU only and GPU enhanced). CPU only is the default and is exact. GPU enhanced needs --features gpu and MOVA_EXEC=gpu. A GPU stage is exact (gated byte for byte) or an approximant (its agreement measured and reported).
• Licences of data (Data and model licences):
  • our own code, models and data: the Elamur Cosmic License (Elamur Cosmic License);
  • NC and research-only data: only for tests and measurements, never in what is built;
  • CC BY-SA: as little as possible. Quote it, test on it, learn from it, but never carry it or modify it into a release.
• Data lives outside the repo, in $MOVA_DATA, and so do run results. The repo holds code, rules, seeds and small data.
• Parallel agents work in their own worktrees. Use git worktree add <scratch>/<task> -b <task> main, with a separate CARGO_TARGET_DIR. Crates depend on en by path, so an edit in a shared copy breaks the others' builds. The main instance merges.
• Commits go in logical steps, after ./secret-guard.sh. It looks for API-key patterns in the staged diff and, if you have one, for the values listed in local.md: an optional, untracked, local file of strings that must never be committed. Contributors may create their own; it is never committed. Its paths: section lists the paths of your machine: the guard refuses a line that adds one outside tera/ (name data through the variables of Mova, «Paths»). Before merging a branch, run ./secret-guard.sh --range main..<branch>: a merge commit does not pass through the guard. Its negative control: tests/secret-guard-control.sh.
• Knowledge goes into the repo (docs/, crate READMEs). Keep docs plain: what it is, how it works, how to run it, measurements, limits.
• Nothing private in the repo. No private paths, personal names or secrets in code, docs or data.

Local extensions: if tera/AGENTS.md exists in your checkout, read it too; it takes precedence, and tera/ is then the only place for private material, which is never published. Public checkouts do not have it.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/AGENTS/