📘 Release 0.5 — notes

Released 2026-11-05. A number on this page is a measurement only once it carries its date. The full list of changes is under "0.5" in Changelog.

0.5 is the release of models of the code (groupants checked by mova-modelcheck, executable Rust approximants), a school student that learns into a saved state and is examined from it in a fresh process, the text–frames–world round trip, parser diagnostics and GPU parsing, translation and Ancient Hebrew parsing, and the animals, the workers built on Mova; its en.bin is retrained on held-out-filtered data with the redacted in-code dictionary.

What is new since 0.4

The details are in Changelog.

• Models of the code. Every Rust source has a groupant: a <file>.groupant/ sidecar with its models and typed obernant families, checked by mova-modelcheck (groupants). Executable Rust approximants (mova-approx) run with 1%, 2%, 5% and 10% error caps (executable approximants). The mova-algorithms library holds deterministic algorithms with their models (algorithms), and mova-correspondence binds them to Mova's sources (correspondences).
• The school student. New reading steps; world knowledge diagnose sorts failed questions by the failing step (diagnosing failures); four controls in every course report must go red (controls); evaluation runs in a fresh process from a saved state (retention); mova-enrich improve is a bounded self-study step (Mova-driven improvement).
• Frames and the round trip. Text → tree → frames → world and back, with a byte-for-byte gate (the round trip, tools/roundtrip).
• Parser diagnostics and GPU parsing. en parse-diag and en construct-eval (en); gparse with fused attention softmax (gparse, GPU stages).
• Translation and ancient languages. translate (tools/translate), the Ancient Hebrew parser (tools/hbo) and the MACULA trees (tools/macula); yom-hadin (yom-hadin).
• The animals. Fish, Crab with its deputy planner, Eel, the Whales, the Sea Spider (MovaBot), the Nautilus and the Sea Turtle; the Dolphin consultants and the per-animal terminals; custom modules through mova-module-overlay (animals). The SLM's training teacher is Codex CLI (teacher transport).

Compared with 0.4

The 0.5 column was measured with the release candidate en.bin bff784ac… before the dictionary redaction; the shipped 2483516a… differs only by that redaction and was not run on held-out sets (see below). The same commands as in Tests, on a CPU with the workspace build. Held-out tests are run once, at release. 0.4 values are from 0.4 notes (2026-10-16).

Build: main 9057160b1, cargo build --release --locked, CPU only, no large-model calls. Every held-out set (bAbI test and its contrast set, the three StepGame test sets, SpartQA-Human, the SVAMP test for both solvers, the tales and EWT tests) was read once.

This table was measured with en.bin bff784ac…, before the redaction; the shipped en.bin is 2483516a…. After the held-out run, 415 lemmas of the in-code dictionary that were personal tokens (e-mail addresses, links, @handles, telephone numbers from the EWT e-mail and web genres) became <redacted-N> placeholders (Data and model licences); organisation domain names seen as words (such as Newsday.com) are not personal data and stay. That changes the dictionary fingerprint, so en.bin was retrained with the same recipe on the same filtered files; two trainings gave the same bytes, 2483516a26a8e17d53c7142df919289cb2a061b37b7b78ae07ae1bba7d41564d, and that model is the one in mova-model-0.5.tar.gz. The held-out sets were not read again. The two models differ only by the redaction; on development sets and the school (2026-11-05, EN_LEXEXT=1, the two builds differing only by the dictionary):

The two models choose the same option on 93.3% of the 65 450 trace rows of the 25 courses. Two runs of the shipped model over the 26 courses gave byte-identical traces and reports, and two runs of the medium bench identical traces.

Notes:

• The parser model is the cause of the bAbI changes. The same 0.5 code with the 0.4 model, on the bAbI train stories: task 16 95.5%, task 18 92.5%; with the 0.5 model: task 16 77.0%, task 18 100%. The 0.5 model reads Does as an auxiliary in "Does the box of chocolates fit in the box?", so task 18 is solved; in task 16 (basic induction) it loses colours such as "Lily is gray" in a story ("no rhino without an attribute"). Task 2 of the contrast set (94.8%) is just under the bar.
• The book's way in cv fell by 6 (1116 → 1110) with the model. With the 0.4 model the same code gives 1116. The school's SVAMP test fell by 1 (Subtraction 87 → 86); it was not rerun with the other model, as it is held out.
• The math solver gained 6 on SVAMP (193 → 199); code and model both changed, and the cause is not isolated. LAS rose on both tests (tales +0.94, EWT +0.60): no sentence of either test is in the 0.5 model's training text (checked by text).
• The 0.5 model is trained on held-out-filtered data. The 0.4 en.bin had the school's test-lesson and dev sentences in its training data, so the 0.4 course score was read with a parser that had seen those sentences. The 0.5 model is trained on EWT train, ESL Spoken train and the school's agreed trees after the held-out filter (training.md), so the course numbers below are not flattered that way.
• The courses. The 25 courses of 0.4 now hold 49 693 test questions: the books course grew from 1 673 to 24 324 questions with the books' chapters, and the books pull the sum down. On the other 24 courses the question sets are close but not identical (25 319 → 25 369 questions; biology 4 581 → 4 631): right answers rise from 11 100 to 12 250 (43.8% → 48.3%), with every control red, despite the cleaner parser. This is not a matched-question comparison, and no bound on the gain over the shared questions is claimed.
• Determinism. Two runs of all 26 courses gave byte-identical traces and reports; two runs of school actors cv and of the bAbI train stories gave identical output; the retained exam run twice from the same saved state gave identical answers.
• The do-support issue of 0.4 is fixed in the held-out run (bAbI task 18 100%), measured with en.bin bff784ac… before the dictionary redaction; the shipped 2483516a… was not run on held-out sets (it picks the same option on 93.3% of course trace rows). The give away verb class is still missing: question_numbers_must_be_used stays ignored.

GPU mode

The stages and their kinds are described in gpu-exec.md.

Both rows were measured with the files of the published mova-model-0.5.tar.gz (SHA-256 bd30339755ad2fcf26857593c019c1b5adfd0eb4eb99ca39b9f12ff6b152c5c0), named explicitly (WORLD_EN_MODEL, MOVA_GPARSE_MODEL): en.bin 2483516a26a8e17d53c7142df919289cb2a061b37b7b78ae07ae1bba7d41564d (CPU column, and the tokens, UPOS, features and lemmas in GPU mode) and gparse/model.safetensors f32dc064999be83b062d02a284c24c18936cb7980532ea4189df15f538f3f373 (GPU column), on an RTX 4000 SFF Ada. Build: commit 17ee53a82e4c7cc3d8d8950d465667e8334f4b15 (the dictionary redaction), world CPU (SHA-256 43346bd2efbc9208f47da4e473e475f7b80a29ea558f308bacddb8e358c2f4c4) and world --features gpu (a1095a06f8cec10c9b97e3e0b36112b323dc87734c8bab702c8fd16486dd190a); the candle CUDA kernels (PTX) were compiled on the same host on 2026-10-22 with CUDA 12.4.131 before the system compiler became GCC 14.2. A clean GPU rebuild from source on that host is blocked (GCC 14 / CUDA 12.4, known issues), so these GPU numbers are repeatable with these binaries but not yet reproduced from a clean build. The fused and the GP_FUSED_SOFTMAX=0 runs were each made twice: all four gave byte-identical traces and reports. With GCC 14 the CUDA 12.4 compiler refuses to build candle's kernels; the GPU build reused kernels compiled earlier (known issues).

Known issues

• bAbI task 16 (basic induction) fell to 74.7% in the held-out run with en.bin bff784ac… (95.3% in 0.4); the same code with the 0.4 model keeps it. The shipped 2483516a… differs only by the dictionary redaction and was not run on the held-out set. Not repaired for this release: the bAbI test is held out.
• No verb class for give away (open from 0.4): question_numbers_must_be_used stays ignored. The do-support issue of 0.4 is fixed in the held-out run with bff784ac… (not re-run with the shipped 2483516a…).
• Unit tests that need the environment (2026-11-05, cargo test --workspace --exclude gparse: 2 601 passed, 15 failed, 12 ignored): the db tests need the DuckDB CLI (DUCKDB) and the wasi-host tests a mlab.wasm built for wasm32-wasip1. The other failures of that run are fixed in the released code: world/tests/groupant_retention.rs reseals its retained states with the dictionary field, roundtrip's lossy_rows_and_before_domain uses its own temporary folder without WAVE32_TEMP, and world's ideas::isolated_tsv_update_and_rank skips without MOVA_HOT_ROOT. The stale lock files of tools/agent-mova and tools/agent-runner found in the same run are regenerated; every standalone workspace resolves with --locked.
• gparse does not build with GCC 14 and CUDA 12.4 (nvcc refuses the host compiler), so --workspace with gparse fails to compile on such a host; the GPU build reused already compiled kernels.
• Resolved: the in-code dictionary's licence is stated (operator's decision, 2026-11-05). en/data/lemmas.tsv and forms.tsv hold forms, lemmas and counts from UD English EWT and are compiled into every binary; the dictionary is distributed under CC BY-SA 4.0, with every source cited in Licence and sources of en/data: the in-code English dictionary (en licence prints it). The citations live outside the compiled blob, so the dictionary fingerprint of en.bin 2483516a… is unchanged (a unit test pins it). Mova's code alone is under the Elamur Cosmic License; its data, the dictionaries included, is under CC BY-SA 4.0 (Data and model licences).
• The Fish terminal crate is still named peta-fish-terminal, and its session, option and report names keep the peta-fish prefix as fixed protocol identifiers (animals).
• The school is a separate archive. The courses, lessons and school dictionaries are not in the repository: mova fetch 0.5 school installs school-0.5.tar.gz into school (MOVA_SCHOOL_DATA picks another folder). Without it the courses are absent, the layer builtin:commands is empty and the register of euphemisms filters nothing. The pupils' dictionary (world/data/dictionary-improved.tsv) stays in the repository, compiled in.
• The school dictionaries are data only. The tools that build them are internal, so they cannot be rebuilt from this repository.
• The peta and worlds features of world add nothing in a public checkout. The legacy manuals they compile in are not in the public tree; without them world/build.rs warns and builds with no manuals. The default build does not use them.
• The worker tools are Unix only. Crab, Octopus, Eel, Whale, the Fish crates and the agent tools use Unix sockets, file modes and libc.

Not carried on

• The 15% and 25% approximant profiles (see "Removed" in Changelog).
• The 0.4 model en.bin (it stays in the 0.4 download).

Downloads

The files that come with 0.5 beside the code, published on 2026-11-05. Each is one .tar.gz, served over HTTPS from https://research.elamur.ai/download/0/0.5/ and by BitTorrent; every torrent has the HTTPS file as its web seed, and our seeder serves them all. Sizes, SHA-256, licences, .torrent files, magnet links and what each archive holds: Mova 0.5 — the release files (for programs: 0.5-files.json). ./install.sh runs mova fetch 0.5, which downloads and checks the model.

en.bin of 0.5 has SHA-256 2483516a26a8e17d53c7142df919289cb2a061b37b7b78ae07ae1bba7d41564d (the archive bd303397…); it is trained on data from which every held-out sentence is removed, and two trainings give the same bytes. The 0.4 files stay at https://research.elamur.ai/download/0/0.4/.

Links

• Changes: Changelog
• Previous notes: 0.4 notes, 0.3 notes
• How to reproduce: Tests
• Licences: Licence files, Data and model licences
• Downloads: above, Mova 0.5 — the release files; release page https://mova.elamur.ai/free/en/releases/0.5/

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/releases/0.5/