Changelog

What changed in each release of Mova. Measurements are in the README and in each release's notes.

Unreleased

yetzirah: Sefer Yetzirah as automata. It holds the letters, their classes and places, the 231 gates, the six seals and the houses of stones, in two classical recensions (standard, Gra) checked quote by quote against the public-domain texts, plus a third reading of ours that holds both where they differ.
The Judaism dictionary (world dictionary judaism, data/dictionary-schools/judaism.tsv): a layer of 6 612 records on the primary school dictionary. It covers every word of the English Tanakh (JPS 1917: the Torah, the Prophets, the Writings) and the English Sefer Yetzirah. For each word, the sense in the Tanakh and Jewish tradition and the tradition's interpretation come first (written by Sonnet), then the corpus sentences with their verse, then the primary examples. requests --update grows it: new words are written, and an entry is reviewed only against sentences it has not seen (the teacher keeps it or rewrites it); the Tanakh round (2026-10-18) added 1 983 words and rewrote 1 238 entries.
Ancient Hebrew Torah (tools/macula): MACULA Hebrew trees into dependency trees (UAS 86.2 / LAS 73.8 against UD PTNK test). It also builds an English dictionary of the Torah's words (BDB 1906, OSHB) with examples in layers: Hebrew, word for word, JPS 1917.
cycle CYCLE_TEACHER: the base teacher of a round can be Haiku (default), Sonnet (Opus reviews above it) or Opus (no cascade); for religious texts, Sonnet 80% / Opus 20%, or Opus alone for hard ones.

The tree reader for word problems (world treeread, school_tree::read_tree): told numbers, actors, verbs, receivers and comparisons are read off the dependency tree into the maths student's world.
Two execution modes (Execution modes: CPU only and GPU enhanced): CPU only (the default, exact) and GPU enhanced (--features gpu, MOVA_EXEC=gpu), where each GPU stage is exact or a measured approximant.
induce and selfplay on the GPU: break counting as an exact integer CUDA kernel, with rules and reports identical to the CPU mode.
The GPU parser (gparse): a transformer trained from scratch on the GPU, also usable as a library.
The schools: primary-school knowledge courses with lessons, exercises, tests and exams, and the pupils' explanatory dictionary.
install.sh: downloads the model en.bin into models with a SHA-256 check.
Corpus 0.4 (https://mova.elamur.ai/free/en/releases/0.4/), under CC BY-SA 4.0 as a whole: the school's books, the agreed school treebank, silver and bronze trees, the pupils' dictionary, maths scripts, contrast sets.
This changelog, a Limits section and a licences table in the README.
One Cargo workspace at the root: cargo build --release builds every CPU crate into target/release/ with no CUDA (the GPU parser on request), with one Cargo.lock (the same dependency versions as the per-crate locks of 0.3).

One model, en.bin: trained on UD English EWT, ESL Spoken and the agreed school trees; every default points at en.bin. The model is not in git.
Licences: the code moved from MIT OR Apache-2.0 to the Elamur Cosmic License (Elamur Cosmic License, licence/); the downloadable models are under CC BY-SA 4.0.
Docs, comments and READMEs in English throughout; private paths replaced by $MOVA_DATA.
Internal tooling and notes (the site server, the sync and paper tools, the coder, the old story worlds) are no longer in this repository.
Measured (2026-10-16): LAS on the tales test 77.38 → 79.69 and on EWT test 79.86 → 80.68; SVAMP (math solver) 64.0% → 64.3%; bAbI 99.8% → 99.4% and its contrast set 99.7% → 99.4% (see Known issues); StepGame and SpartQA unchanged.
The two unfinished courses (programming, philosophy) are no longer part of the release.

en.bin misreads do-support questions with a long subject ("Does the box of chocolates fit in the box?"): bAbI task 18 92.3%, task 16 95.3%. The same code with the 0.3 model gives 99.8% on bAbI.
The maths reader has no verb class for give away; the unit test question_numbers_must_be_used is ignored until 0.5.

The 0.3 model ud-ewt-eslspok.bin is not carried on.
The maths student's old window-of-words reader; the tree reader is the only reader.
The story worlds without a teacher's check moved out of the default build (the schools replace them).

Self-retraining with states (Self-play: the SLM modifies itself cheaply, tests, reports, resets): a tree of delta states with rollback, a brain and decision letters per state, and the family training loop on a Unix socket with live control.
Learning on big texts: samples from the tree store judged by the absurdity matrix, UD grammar constraints and a teacher check; a negative list of rejected rules.
The absurdity matrix (Absurdity matrix — the central common-sense tool): 405 verbs × 601 nouns scored 0–4, plus a fairy-tale layer.
Induced tree-repair rules (The SLM derives its own tree rules) and trees by logical derivation (Trees by logical derivation).
StepGame corrected and contrast sets: a reader of new phrasings next to the learned perceptron.
Expressions with their real meanings (Expressions with their real meanings — the pragmatics lens) and the tree store (world store).
Corpus 0.3 on research.elamur.ai.

SpartQA-Human yes/no 60.1% → 75.5%, relation questions 50.6% → 63.6%; SVAMP 59.7% → 64.0%; LAS on EWT test 79.82 → 79.86.
The first verified self-training gain shipped: nsubj/obj → obl when the word has a preposition.

The first free release.

The English UD parser (en), pragmatics (prag), story worlds (world), the maths solver (math) and the compiled global knowledge (global).
bAbI 99.8%, StepGame 96.3%, SpartQA-Human yes/no 60.1%, SVAMP 59.7%, LAS on EWT test 79.82.
Corpus 0.2 on research.elamur.ai.
Licence: MIT OR Apache-2.0.