Release CAEL 0.3.8: code-only update on the 0.3.7 model bundle (AI-assisted)

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+cael-0.3.7.cael filter=lfs diff=lfs merge=lfs -text

CAEL-0.3.8.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0d1771d54e9e5b77e84462cc00411e5cadede076ae4709fe06440c9e535dd6e
+size 21562557

CHANGELOG.md

@@ -0,0 +1,45 @@
+# Changelog
+
+## 0.3.8 (release candidate, not yet published)
+
+Code-only release. It loads the unchanged published `cael-0.3.7.cael` model bundle
+(same SHA-256 checksum); no weights or bundle files changed. AI-assisted.
+
+### Fixed
+
+- Personal facts: CAEL no longer invents a personal fact it was never told (for example a
+  birthplace). It answers from what the user said in the conversation, or says it has not
+  been told.
+- Listening requests: an emotional message that asks for no advice and no questions now gets
+  a short reflection without advice or questions, for more phrasings than before.
+- Word problems: a checked quantity graph handles more multi-step quantity problems
+  (receive, give away, buy, share). The older exact inventory helper still answers first
+  when it can, so its worked steps and multi-turn "Add:" state are kept. A leading
+  "and"/"then" in a step is handled; "to" is no longer taken as part of a quantity noun.
+- Linear equations such as `Solve 6*x + 5 = 47` are solved exactly.
+- Python generation: the requested function name and signature are kept; a single function
+  stub is checked against the request; prompts that contain real code plus extra steps are
+  left to the task and data tools instead of being asked to clarify.
+- Code execution: generated specs and recurrences run only inside the restricted Python VM
+  (no native `eval`). The VM gained a bounded `dict.fromkeys`.
+
+### Evaluation (read-only, no tuning on evaluation items)
+
+- 80-item read-only replay, raw automatic grading: 59/80, against 54/80 recorded for 0.3.7
+  (memory 17/20, conversation 10/20, Python 16/20, controls 16/20). Seven replies changed:
+  memory-20, conversation-04, python-15 and controls-05 now pass; controls-16 now passes the
+  automatic check (it had been accepted in review before); python-05 and python-19 still fail.
+  Memory-20 is the item from the public bug report, so it is no longer a clean held-out item.
+  Review of the replies was done by AI coding agents (unblinded).
+- Public 128-question comparison: math 0/32 and Python 1/32, the same as 0.3.7. CAEL does
+  not pass the competitive math gate and 0.3.8 makes no math improvement claim.
+- GSM8K train split, 400-question development check (informational only): 1 correct, 17 wrong, 382 abstained. Coverage is very low and the 17 wrong answers come from routes that should have abstained.
+- The 400-question comparison is unchanged: Qwen3.5 leads clearly (181/400 original phrase
+  passes against CAEL 121/400; Qwen3 116/400). Qwen is also better at open conversation.
+
+### Known limitations
+
+- Bundle builds still name the bundle after the code version; 0.3.8 ships the 0.3.7 bundle.
+- The restricted VM refuses `math.*` expressions.
+- Word-problem coverage is still low; some routes give wrong numbers instead of abstaining.
+- Python-19 still uses an import although the prompt asks for no imports.

LICENSE

@@ -0,0 +1,70 @@
+CAEL NON-COMMERCIAL LICENSE (Personal and Research Use Only)
+
+Copyright (c) 2026 Vovala14. All rights reserved except as granted below.
+
+This license applies to CAEL (Cellular Automaton Event Language) and to the
+code, models, weights, prompts, data, documentation and other files in this
+repository and in CAEL model bundles that are owned by the copyright holder
+(the "Software"). Third-party components are covered by section 8 instead.
+
+1. GRANT OF PERMISSION
+   Permission is granted, free of charge, to any person obtaining a copy of
+   the Software to use, copy, modify, and run it, solely for:
+   (a) Personal Use: private, non-commercial use by an individual for their
+       own learning, hobby, or experimentation; and
+   (b) Research Use: non-commercial academic, scientific, or educational
+       research, and publication of the results of that research.
+
+2. COMMERCIAL USE REQUIRES A SEPARATE LICENSE
+   You may NOT use the Software, or anything derived from it, for any
+   commercial purpose without a separate written license from the copyright
+   holder. Commercial purposes include, without limitation:
+   (a) selling, licensing, sublicensing, renting, or otherwise charging for
+       the Software or derivatives, or access to them;
+   (b) using the Software in, or to provide, a product or service that is
+       sold, or that generates revenue, directly or indirectly (including
+       through advertising, subscriptions, or paid APIs);
+   (c) using the Software in the operations of a for-profit business or
+       for any other commercial advantage; and
+   (d) using the Software or its outputs to train, fine-tune, or distill
+       models or systems that are used commercially.
+   The copyright holder keeps all commercial rights and may use and license
+   the Software commercially.
+
+3. REDISTRIBUTION
+   You may share copies or modified versions only for non-commercial
+   purposes, and only if you (a) include this license and the copyright
+   notice with them, (b) clearly mark any changes you made, and (c) license
+   your distribution of the Software under these same terms. You may not
+   remove or alter this license or the copyright notice.
+
+4. ATTRIBUTION
+   Any research publication or public work that uses the Software must
+   credit "CAEL by Vovala14" and link to https://github.com/Vovala14/CAEL .
+
+5. OWNERSHIP
+   The Software is licensed, not sold. The copyright holder retains all
+   rights not expressly granted here. No trademark rights are granted.
+
+6. TERMINATION
+   This license terminates automatically if you breach it. On termination
+   you must stop using the Software and delete your copies.
+
+7. NO WARRANTY AND LIMITATION OF LIABILITY
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+   OR IMPLIED, INCLUDING BUT NOT LIMITED TO WARRANTIES OF MERCHANTABILITY,
+   FITNESS FOR A PARTICULAR PURPOSE, AND NONINFRINGEMENT. IN NO EVENT SHALL
+   THE COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY,
+   WHETHER IN CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN
+   CONNECTION WITH THE SOFTWARE OR ITS USE.
+
+8. THIRD-PARTY COMPONENTS
+   Third-party data, software and graphics included in or used to build CAEL
+   keep their own licenses, listed in licenses/, in each model folder's
+   SOURCES.md / TRAINING_NOTICES.md and in the bundle. This license does not
+   replace or restrict them. In particular, components derived from Creative
+   Commons Attribution-ShareAlike material (for example the Wikipedia,
+   Stack Exchange, Dolly, Wikibooks, Wikivoyage and Universal Dependencies
+   derived knowledge and count tables) are shared under their CC BY-SA terms,
+   which also allow commercial use. CAEL only uses third-party material whose
+   license allows commercial use.

NOTICE.md

@@ -0,0 +1,47 @@
+# Provenance and separation
+
+CAEL (Cellular Automaton Event Language) is a separate model family owned by Vovala14. Version0.1 derives from the locally preserved constrained symbolic event-CA experiment of2026-10-13. It is not the native MICA architecture or a MICA foundation-model release.
+
+The original experiment is retained unchanged outside this repository. Earlier CAEL releases (up to 0.2.9 / 0.3) carried that experiment's licence text verbatim, which named MICA Ember / MICA Flame. CAEL and MICA have the same copyright holder (Vovala14), who now licenses CAEL under its own CAEL Non-Commercial License (LICENSE): free for personal and research use, with commercial use reserved to the copyright holder. CAEL is a separate model family and contains no MICA code, weights or outputs. CAEL is not released under an OSI open-source license.
+
+## Commercial-use policy for third-party material
+
+CAEL only uses third-party data, software and graphics whose licence allows commercial use (MIT, Apache, BSD, CC BY, CC BY-SA, WordNet, Unicode, CreativeML OpenRAIL-M). Removed in 0.3.1: the 0.2.1 small-talk reply models (`chat/model_v1.json.gz`, `chat/model_v2.json.gz`), which were fitted on EmpatheticDialogues (CC BY-NC 4.0). The ARC Corpus is never used at runtime (its terms forbid redistribution); a local research reproduction can opt in only with `CAEL_RESEARCH_ARC=1`.
+
+Changes: standalone portable training/evaluation, CAEL naming, cumulative event-memory training, explicit multi-event API, stricter supported-question handling, reports and public packaging. Runtime includes handwritten parsing and slot formatting plus learned cellular event rules, plans and a template-like regular grammar. No neural model is used for training or inference; the curriculum is expert-authored synthetic data.
+
+## CAEL 0.2 emoji model and data attribution
+- models/cael-emoji-64-v2 was trained on Noto Emoji (Copyright Google Inc., Apache License 2.0, github.com/googlefonts/noto-emoji @ e20cbc2b, 2D/png/128)
+  and Blobmoji (C1710 and contributors, based on Noto Emoji, Apache License 2.0, github.com/C1710/blobmoji @ 7dd14d2b, SVGs rendered to 128x128 PNG).
+  The model stores learned counts, not the images. Licences in licenses/.
+- Captions: Unicode emoji-test.txt short names + Unicode CLDR English annotations (Unicode License).
+- Transparent pixels were composited on white before training.
+
+## CAEL 0.2.1 changes and synthetic-image teacher
+
+0.2.1 adds a shared application, one integrity-checked model bundle, portable
+legacy chat fallback, exact arithmetic, the current v7 emoji engine, sanitized
+v15 aggregate portrait parameters, and a new categorical general-image student.
+The preceding description of expert-authored training and no neural model
+applies to Story 0.1, not the new synthetic-image training pipeline.
+
+The general-image student is trained on outputs generated locally by neural
+Stable Diffusion 1.5 through stable-diffusion.cpp. It is treated as a derivative
+under CreativeML OpenRAIL-M. Original diffusion/text-encoder weights are replaced
+by subject-specific registered 64x64 statistical appearance/deformation modes.
+Use and redistribution must comply with the included teacher license, paragraph
+5 and Attachment A, in addition to the CAEL Non-Commercial License for owned
+modifications. Full teacher license: licenses/LICENSE-stable-diffusion-v1-5.txt.
+Primary sources: https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5
+and https://github.com/CompVis/stable-diffusion/blob/main/LICENSE.
+
+stable-diffusion.cpp is a separate locally installed MIT-licensed teacher tool:
+https://github.com/leejet/stable-diffusion.cpp. Neither its binary nor the neural
+teacher checkpoint is redistributed in this release. Runtime student sampling
+is non-neural, but its neural teacher's contribution is explicit.
+
+v7 emoji data sources remain Noto Emoji and Blobmoji, with the same Apache and
+Unicode attribution above. Public fitting references are included for numerical
+rejection and silhouette statistics. Private portrait photos and alignments,
+endpoint credentials and private chat logs are not included.
+

README.md

@@ -0,0 +1,126 @@
+---
+language:
+- en
+license: other
+license_name: cael-non-commercial
+license_link: https://huggingface.co/vynly/CAEL-0.3.8/blob/main/LICENSE
+pipeline_tag: text-generation
+tags:
+- cael
+- non-neural
+- cellular-automata
+- symbolic-reasoning
+- conversation-memory
+- code-generation
+inference: false
+---
+
+# CAEL 0.3.8
+
+A CPU assistant built from non-neural symbolic, count and automaton engines.
+It is a custom Python runtime; Qwen and Transformers are not part of CAEL.
+CAEL is source-available, non-commercial software (see LICENSE).
+
+0.3.8 is a code-only update. It reuses the unchanged published 0.3.7 model
+bundle (cael-0.3.7.cael, same checksum); see CHANGELOG.md.
+
+Picture and portrait generator weights are excluded. Chat, helper, math,
+stories, prompt writing, retrieval and emoji remain available. Image requests
+can produce a prompt for an external image tool.
+
+## Start
+
+Use Python3.11 or later. Extract the source archive or clone this repository,
+then run in its folder:
+
+```bash
+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. Windows SETUP.bat and RUN.bat provide the same flow.
+The checksum-verified no-images resource download is about563MB; source and
+resource archives are separate. Existing local models/selections are preserved
+by the installer. The licence is unchanged; see LICENSE and NOTICE.md.
+
+## Changes in 0.3.8
+
+- Personal facts: a question about your own details (birthplace, address and
+  similar) that you never stated is answered with "you have not told me",
+  instead of a fact borrowed from retrieval or another person.
+- Requests to just listen ("no advice", "don't ask me questions", and
+  paraphrases) are kept on every conversational route, not only some phrasings.
+- Word problems: a checked quantity-event graph, exact fractions/decimals and
+  stricter clause handling. Unsupported clauses lead to a clarifying question.
+  The older inventory helper still answers first when it is exact.
+- Python: requested function names and declared signatures are kept, and
+  rules that cannot be resolved lead to a question, not to simpler code.
+- Safety: arithmetic-expression and recurrence checks no longer use Python's
+  native eval. They run in the bounded AST VM or a closed +, -, * evaluator.
+- The model weights and bundle are unchanged.
+
+## Changes in 0.3.7
+
+- Session-local explicit facts/preferences, corrections, conflict clarification,
+  forgetting and provenance. Quoted/hypothetical statements are treated cautiously.
+- A larger authored vocabulary for mixed/negated feelings, speaker attribution,
+  bereavement and requests to listen. These are bounded rules, not human experience.
+- Natural-language ordered numeric/text/record workflows compose existing Python
+  operations; simple list/array chunk contracts are restored. Generated functions
+  are checked in a bounded AST VM. Unsupported clauses need clarification.
+- The learned model weights, protected choice/word-math solvers and CA kernels
+  are unchanged. These upgrades are implementation changes.
+
+## Named local confirmation
+
+CAEL-CMP-R2-20261028 contains 80 locally authored synthetic items, 20 per category.
+Every full reply was reviewed by AI coding agents (unblinded), including apparent
+automatic passes and code.
+The source, inputs and primary grading were frozen before final answers.
+
+| System | Memory | Conversation | Python | Controls | Reviewed total |
+|---|---:|---:|---:|---:|---:|
+| CAEL 0.3.7 | 16/20 | 10/20 | 15/20 | 16/20 | 57/80 |
+| Qwen3-0.6B | 8/20 | 13/20 | 6/20 | 16/20 | 43/80 |
+| Qwen3.5-0.8B | 12/20 | 14/20 | 3/20 | 12/20 | 41/80 |
+
+Functional Python: CAEL15/18, Qwen3 6/18, Qwen3.5 3/18. Original automatic totals
+were54,43,42; full-review totals differ because regex matches can miss correct
+answers or accept incorrect explanations. All raw replies/flags and every
+adjudication are included under benchmarks/cael_cmp_r2_20261028/.
+
+0.3.8 offline replay of the same 80 frozen inputs (automatic grader, no
+re-review, no Qwen rerun): 59/80 automatic, versus 54/80 for the recorded 0.3.7
+replies (memory 17, conversation 10, Python 16, controls 16). Seven replies
+changed, listed in CHANGELOG.md. One of them (an invented birthplace) is also the
+bug in a public report fixed in this release, so that item is not a clean
+held-out test any more. The table above remains the 0.3.7 reviewed result.
+
+This is an authored family-overlap test with unblinded review by AI coding
+agents, not an independent leaderboard or evidence of general superiority. Qwen is stronger
+on the conversation category here. On an older, broader 400-question test,
+Qwen3.5 leads clearly: CAEL 121/400 versus Qwen3.5 181/400 original phrase
+passes (Qwen3 116/400), with phrase-grading flaws disclosed. The targeted result
+must not erase that limitation.
+
+CAEL's measured full-portable process peak was558.648MiB in this shared-host
+run. Qwen3 float32/eager peaked5896.336MiB; Qwen3.5 bfloat16/SDPA peaked2668.309MiB.
+These are different runtimes/precisions, not an equal-size or optimized quantized
+comparison. Disk size, model bytes, load time and median latency are separate
+quantities recorded in the report. The earlier167MiB source-only run is not
+this portable-build measurement.
+
+## Remaining weaknesses
+
+CAEL still misses factual/emotional paraphrases and fails tracing, leftmost
+binary search under a no-imports rule, and much reasoning. It solves few
+ordinary grade-school word problems: on a 400-problem GSM8K training sample
+(a development check, not a release benchmark) it gave very few correct answers
+and asked for clarification on most. Qwen is better at open conversation.
+It does not understand arbitrary Python or replace a general LLM.
+Do not enter personal information into shared demos. This release does not
+change any running Vynly site or shared worker.
+
+Read CHANGELOG.md, RELEASE_0_3_7.md and benchmarks/cael_cmp_r2_20261028/REPRODUCE.md for exact
+settings, uncertainty, disclosures and the offline CAEL reply-regression replay.

SHA256SUMS.json

@@ -0,0 +1,10 @@
+{
+  "CAEL-0.3.8.zip": "c0d1771d54e9e5b77e84462cc00411e5cadede076ae4709fe06440c9e535dd6e",
+  "CHANGELOG.md": "6e7476618384caacbeed11890bcb28268f0017d02f4cd733a88c3e44b8d3a050",
+  "LICENSE": "d5d32c0f00aa08ac59843f80cd36b51717c6365585a9ead644ddf0db1f492661",
+  "NOTICE.md": "d12deb61adb80197d3976b97540c103bbfe85e5b9281abb0e94a468bbf6f53dc",
+  "README.md": "e08b22812c038d0e65a9ac4488c7df96211cefcb087ba20560ed201e19d8e9d2",
+  "bundle_config.json": "8aa7534c18a7be3f340834becad6e5d9ebbebec3e69d5a905e975c548fc34e44",
+  "cael-0.3.7.cael": "8efd22d7fa670e7f011d262c33a7e026941f279e71ee2014614d5d4cf5ecf79b",
+  "release_manifest.json": "479f5a095634cecab86f07ea76b4fede9c75c116e50b7e0369f4edd0cd24c900"
+}

bundle_config.json

@@ -0,0 +1,7 @@
+{
+  "version": "0.3.7",
+  "url": "https://github.com/Vovala14/CAEL/releases/download/v0.3.7/cael-0.3.7.cael",
+  "sha256": "8efd22d7fa670e7f011d262c33a7e026941f279e71ee2014614d5d4cf5ecf79b",
+  "bytes": 562714794,
+  "profile": "no-images"
+}

cael-0.3.7.cael

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8efd22d7fa670e7f011d262c33a7e026941f279e71ee2014614d5d4cf5ecf79b
+size 562714794

release_manifest.json

@@ -0,0 +1,64 @@
+{
+  "version": "0.3.8",
+  "profile": "no-images",
+  "source_commit": "7a2322743ec8083eb3eef2b257e99289831735dc",
+  "source_archive": {
+    "file": "CAEL-0.3.8.zip",
+    "bytes": 21562557,
+    "sha256": "c0d1771d54e9e5b77e84462cc00411e5cadede076ae4709fe06440c9e535dd6e"
+  },
+  "resource_archive": {
+    "version": "0.3.7",
+    "url": "https://github.com/Vovala14/CAEL/releases/download/v0.3.7/cael-0.3.7.cael",
+    "sha256": "8efd22d7fa670e7f011d262c33a7e026941f279e71ee2014614d5d4cf5ecf79b",
+    "bytes": 562714794,
+    "profile": "no-images"
+  },
+  "code_only_release": true,
+  "resource_archive_unchanged_from": "0.3.7",
+  "checks": {
+    "pytest_passed": 1770,
+    "pytest_failed_environment_or_preexisting": 4,
+    "pytest_failed_on_v0_3_7_baseline_too": 4
+  },
+  "replay": {
+    "name": "CAEL-CMP-R2-20261028",
+    "items": 80,
+    "raw_automatic": {
+      "0.3.7_recorded": 54,
+      "0.3.8": 59
+    },
+    "per_category_0.3.8": {
+      "memory": 17,
+      "conversation": 10,
+      "python": 16,
+      "controls": 16
+    },
+    "changed_replies": 7,
+    "review": "AI coding agents, unblinded; not an independent evaluation",
+    "note": "memory-20 comes from a public bug report and is no longer a clean held-out item"
+  },
+  "public128_math_python": {
+    "math": "0/32",
+    "python": "1/32",
+    "same_as_0.3.7": true,
+    "competitive_math_gate_passed": false
+  },
+  "gsm8k_train_dev_400": {
+    "correct": 1,
+    "wrong": 17,
+    "abstain": 382,
+    "informational_only": true
+  },
+  "comparison_400": {
+    "CAEL": 121,
+    "Qwen3.5-0.8B": 181,
+    "Qwen3-0.6B": 116,
+    "note": "Qwen3.5 leads; evidence kept in local research files, not in this archive"
+  },
+  "image_portrait_weights": false,
+  "private_data_included": false,
+  "opponent_weights_or_runtime": false,
+  "licence": "source-available, non-commercial; unchanged",
+  "attribution": "AI-assisted"
+}