CAEL 0.3.6
A CPU assistant combining non-neural symbolic, count and automaton engines in one portable model bundle. Chat, conversation reasoning, exact math, bounded Python helpers, retrieval, stories, prompt writing and emoji. Picture and portrait generation weights are excluded. This is a custom Python runtime, not a Transformers checkpoint.
Install
Download and extract CAEL-0.3.6.zip, then run in its CAEL-0.3.6 folder:
python -m pip install -r requirements-unified.txt
python -m cael_unified --download
python -m cael_unified --ui
Open http://127.0.0.1:8765. The download command fetches and verifies the approximately 563 MB model archive from
GitHub. Alternatively download cael-0.3.6.cael here, put it in the extracted source folder's
models/ directory, and run python -m cael_unified --download to verify/install the existing file.
Fresh installation is recommended; existing local selections and differing model files are preserved.
Changes
Structured conversation records, clause-based context reasoning, corrections and conflicting sources, paraphrased recall and bounded operations over stated values. Spelling lookups use compact memory-mapped integer/hash tables to reduce resident Python objects. The learned model files are unchanged from 0.3.4; the context improvements are code changes. The language architecture is retained, with no neural runtime.
Evaluation and limitations
Saved final runs on the same locally authored, heuristic-graded 400-question context benchmark:
| System | Correct | Accuracy |
|---|---|---|
| CAEL 0.3.5 | 53/400 | 13.25% |
| CAEL 0.3.6 | 122/400 | 30.50% |
| Qwen2.5-0.5B-Instruct (float32) | 102/400 | 25.50% |
| Qwen3-0.6B (float32) | 121/400 | 30.25% |
| SmolLM2-360M-Instruct (float32) | 28/400 | 7.00% |
The one-answer difference from Qwen3 is effectively a tie. These are local conversation tests, not a general capability ranking, external independent evaluation, or a new HellaSwag/Tiny ToM leaderboard score. CAEL still misses 278/400; references, planning and unsupported-answer handling remain weak. LMs used CPU/eight threads, greedy decoding and 64 output tokens; CAEL records one thread. Qwen3 used its non-thinking template. The scorer is phrase based, so matches are not a substitute for human semantic review.
Release checks: 172 tests plus 129 subtests passed in the focused text regression selection. All 97 bundle members were checksum-verified; 73 packaged runtime files match the local models. A clean-install text-only smoke run is recorded in the release manifest. No private source photos or private conversations were opened.
See release notes and the source archive for protocol, category breakdowns, frozen code, synthetic questions, saved outputs and reproduction instructions. Earlier development-informed scores must not be presented as independent performance. Full source commit: 41ffa73543a84734471e9515a636efc573d406b4.
Files and licence
CAEL-0.3.6.zip: source from the pinned Git commit, with per-file checksums.cael-0.3.6.cael: portable model bundle, no picture/portrait models.SHA256SUMS.json: public file integrity manifest.bundle_config.json: model download URL, size and checksum.
CAEL Non-Commercial License: free for personal and research use; commercial use requires a separate licence. Third-party components retain their own terms and attribution; see NOTICE.md and the licence files in the source/model archives.