📘 Documentation

Mova is a Hybrid Combinatorial–Operational System with a Combinatorial Small Language Model (SLM) at its core. These pages describe how it is built and how it works. Start with the hybrid system, then the architecture. Each crate has a README next to its code with the full command reference.

Overview

• Mova as a hybrid system — what Mova is: the three layers (language, combinatorial, operational), what is implemented and what is direction; the octopus mascot.
• The economics of Mova — why Mova changes where intelligence is spent: a persistent orchestrator, the SLM at the margin, large models as consultants, and what to measure.
• Architecture — the pipeline from text to answer, the three levels of logic, the teacher loop, the students, the crates and where things live.
• Context, layers and memory — context, layers and memory: the tale world, the SLM subsystems and what is inferred from context, with the gaps named.

The language layer

• The English layer — the English parser and compressor step by step, the grammar expert system, trees by derivation, and the round trip through the layers.
• Open parsers: witnesses and tests — the open parsers Mova uses as witnesses and tests, their licences and roles.
• The Mova dialect: a UD metadialect, treebanks and converters — the Mova UD dialect, treebanks and converters between conventions.
• Pragmatics: frames — who did what to whom, why, and what came next — frames (who did what to whom, why, and what came next), speech acts, implicature, expressions read by their meaning, the external pragmatics tools.
• seeds/ — the grammar seeds explained, one page per folder.

Worlds, knowledge and the school

• Worlds and knowledge — the knowledge contract (skeleton, important graph, global graph), the absurdity matrix, the nerve graph, RAG.
• The school — the primary school: how courses are made, their format, how the pupil reads and answers, controls, diagnosis, tools, retention in a fresh process, Mova-driven improvement, and the school results.
• Training Mova — how Mova is trained: the teacher transport, the training pipeline, measuring and gates, self-directed learning (induce, self-play), the annotation and storage contract, reviewed promotion, re-annotation.

Workers

• Animals and workers — the Octopus (an LLM that develops Mova, not part of its runtime) and Mova's Rust animals: Crab, Eel, Fish, Whales, Nautilus, Turtle, Sea Spider (and MovaBot), terminals, the Dolphin consultants; their protocols, evidence boundaries and custom modules.

Groupants and mathematical modeling

• Groupants — the groupant: the combinator node of a piece of code, its components, sidecars, tree, registry, coverage, the external reference library and the publication scope.
• Mathematical modeling of code — the models every file and group needs: vidtvorants, typed obernants, reverse reconstruction, the mathematical combiner, executable Rust approximants.
• Correspondences — typed mathematical correspondences between algorithm groupants and Mova sources, and algorithm transfers through explicit adapters.

Execution, data and releases

• Execution modes: CPU only and GPU enhanced — the two execution modes, CPU only and GPU enhanced: stages, checks and reference switches.
• Data, libraries and licences — where data lives, licences, the maths and science libraries, Wiktionary, paper cards, competitions.
• Add-on modules — add-on modules (for example dictionaries under their own licences): install from a URL or a torrent, verify, and load into Mova.
• Changelog — what changed in each release; release notes: 0.6 (draft), 0.5, 0.4, 0.3, all in releases/.
• Elamur Cosmic License and Data and model licences — the Elamur Cosmic License and the licences of data and models.

Elsewhere in the repository

• Tests — unit tests and benchmark commands.
• corpus/open — the open corpus. Papers Mova uses lists the works Mova uses, each linked to its card on Elamur Research; usage.tsv logs where each is applied.
• train — how to retrain Mova for your own domain.
• eval/ — evaluation lists. Ablation reports of tools/value are run records, not pages.
• assets/ — the octopus artwork. mysticism/ — background texts and Gali, the working mode for AI instances that build Mova: Mova's own working philosophy, not tooling; nothing in the build or the checks depends on it.

Writing docs

New documentation extends one of the pages above. A new page needs a new subsystem; per-feature, per-experiment and per-review notes do not become pages. Docs describe how things are and work; the only measurements kept are school results and comparisons with earlier releases. Every relative link must resolve: cargo test -p mova-modelcheck --test doc_links checks files and heading anchors in all public Markdown.

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