vynly/CAEL-0.3

CAEL 0.3 — helpers and local retrieval

CAEL is a hybrid non-neural assistant: counted language models, cellular automata, symbolic reasoning, exact arithmetic, constrained Python helpers, local retrieval, prompt writing and story generation. It uses no neural network, transformer or external language model at inference. This is a custom Python runtime, not a Transformers/AutoModel checkpoint.

Version 0.3 publishes the evaluated 0.2.9-based MC-helper candidate under a new public release name. Learned weights are unchanged; this is a runtime/packaging release, not a newly trained language model. The native count-language checkpoint remains 0.3.2-math-python; that internal component version is distinct from this unified release version.

Download and run

1. Download CAEL-0.3-source.zip and extract it into a new folder.
2. On Windows, run SETUP.bat, then RUN.bat. Setup installs dependencies and downloads the checksum-verified 555.6 MB model bundle from this repository.
3. Open http://127.0.0.1:8765 for the local chat UI.

For terminal use, with Python 3.11/3.12 and dependencies installed:

cd CAEL-0.3
python -m pip install -r requirements-unified.txt
python -m cael_unified --download
python -m cael_unified "Calculate 1/3. Then multiply it by 3" --mode chat
python -m cael_unified --ui

For an offline installation, download cael-0.3.cael, place it in the extracted source folder's models/ directory, and run python -m cael_unified --download to verify and install its members. Dependencies must already be available. SHA256SUMS.json pins the downloadable files; the extracted source ZIP has its own complete file checksum manifest.

The source contains the new choice_helpers.py adapter and matching Assistant.respond integration. Model archive metadata and the staged runtime report version 0.3. Earlier releases and the existing live Vynly worker are not changed by this publication.

What is included

• The unified model archive with the existing language, chat, reasoning, helper, story, prompt, dictionary and Wikipedia knowledge components.
• Exact arithmetic, supported symbolic/state/ordering MC routing, and the existing evidence/calculator-assisted chooser.
• Runtime source, tests, dependency specifications, notices and licenses.
• Benchmark prediction/audit evidence without benchmark question text, option text or gold labels.
• Emoji support is retained. Picture/portrait/general-image generation checkpoints and their active selections are excluded. Picture requests can produce text prompts instead.

The included Wikipedia index is 376,369,152 bytes uncompressed, inside the model bundle. The archive's 95 existing model members retain their previous SHA-256 values. Only bundle version/provenance metadata changed.

Retrieval limitation and ARC licensing correction

The historical full assisted evaluation used a separate 2,610,143,232-byte ARC Corpus SQLite index. Neither that index nor the underlying ARC Corpus is redistributed here. A normal installation uses the included Wikipedia index and available helpers, so it does not reproduce the full assisted benchmark configuration below.

The local original ARC-V1-Feb2018-2/README.txt includes corpus-specific restrictions on commercial use and redistribution, including: “You may not redistribute this dataset to others.” The CC BY-SA label on the ARC question dataset card is not sufficient evidence to relicense the separate science-sentence corpus. Some historical CAEL documents/code comments called the corpus CC BY-SA; this release notice corrects that attribution without changing the frozen solver code.

The ARC component is not automatically downloaded, built or enabled by this package. Any separate acquisition/use requires the publisher's applicable permission and terms; no commercial-use clearance is claimed. The excluded index hash is recorded only for evaluation provenance. We do not claim that the 24.37 assisted result is reproducible using the default download or commercially cleared end to end.

Evaluation

Local results on all 16,428 pinned questions:

Task	Questions	Native count-language	Historical helpers + Wikipedia + ARC
HellaSwag validation	10,042	27.972515%	27.932683%
ARC-Easy test	2,376	26.599327%	62.121212%
ARC-Challenge test	1,172	22.184300%	31.655290%
PIQA validation	1,838	48.911861%	58.650707%
ArithMark-3	1,000	37.800000%	69.500000%
Current Open SLM Intelligence Index		3.306710	24.373331

These are not official accepted leaderboard results. They were obtained with a custom backend, not a claimed upstream LM-eval harness run. Native likelihoods were freshly recomputed. Assisted results are a hash-verified cached differential evaluation of the unchanged fallback system, with 15 fresh full-chooser/retrieval parity checks. The new MC helper branches answered zero of these SLM questions; this release does not claim an SLM score increase.

Benchmarks and prior outcomes were publicly known and used repeatedly; adapters were informed by public NNAI formats. Four labeled ArithMark examples were exposed in earlier format inspection. No benchmark rows or labels were fitted in this release; no hidden-test generalization, comprehensive contamination audit, or independent organizer verification is claimed. No neural-style parameter count has been established for this hybrid system; archive bytes are not a parameter count.

Evidence is in evaluation/. Open SLM PR #256 requests an eligibility/methodology decision for the assisted system. Organizers have not accepted or ranked it.

License, provenance and limitations

Owned code/model materials retain the existing MICA AI Non-Commercial License; this publication does not relicense them. Third-party component terms and notices are retained in the source archive and bundle. Wikipedia content remains under its attributed terms. Public download access is not an OSI-open-source or commercial-use claim.

Base source: Vovala14/CAEL, commit 0055d9863080f0beeaed31978116559914f1be10. Exact evaluated candidate hashes, packaging changes and prior model hash are recorded in release_manifest.json. No private photos, crops, conversations, user credentials, neural teacher weights or unselected local experiments are published.

CAEL has narrow supported symbolic/programming forms and a weak count-language baseline; it does not provide general Python understanding or reliable open-ended reasoning. Its code/tool proposals need review before execution. Runtime use of PyTorch for existing count/tensor-based components does not imply a neural architecture.