vynly/CAEL-0.3.9

CAEL 0.3.9

A custom Python assistant built from non-neural symbolic, count and automaton engines. This release improves conversation follow-ups, explicit state tracking, checked arithmetic/Python rechecks, personal-evidence attribution and bounded source lookup. It does not use Transformers, a neural model or Qwen at runtime.

The 0.3.9 source includes the qualified symbolic dialogue/state tables. The larger public resource archive remains the byte-identical cael-0.3.7.cael; its version is reported separately from the 0.3.9 application. Picture/portrait weights are excluded from this public package. Chat can prepare an image prompt instead.

Run

Use Python 3.11 or later. Extract CAEL-0.3.9.zip, then run in the extracted 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. For a terminal conversation use python -m cael_unified --mode chat. A reset starts a new visible conversation; do not use the shared demo for sensitive information. Public lookup uses free search/page readers and can be disabled with CAEL_WEB_SEARCH=0.

This is a custom runtime, not a Transformers-compatible checkpoint. Both the source archive and the approximately 563 MB resource archive are required. The installer preserves existing local model selections and reports conflicts. Licences and attribution are retained in LICENSE, NOTICE.md and licenses/.

Improvements

• Follow-up requests retain supported task, style, option and user-evidence state. Arithmetic rechecks recalculate the original request; unrelated tasks block stale reuse.
• Complete literal Python examples, including keyword calls, are checked through the existing bounded AST interpreter. The checker policy used in qualification is installed automatically by the public Assistant entry point.
• Finite readers handle exact quantities, explicit owners/units/times, conditional options and dictionary workflows. Unsupported or incomplete contracts clarify.
• Definition comparisons can retain two source-linked definitions. Public search extends beyond Wikipedia; retrieved text is evidence, not proof of truth.
• Request learning is ephemeral by default. Explicit supplied histories retain within-conversation context; unrelated sessions do not share learned rules.

Measured scope

The latest authored conversation study compares this candidate with the preceding conversation R2 development candidate, not the published worker or Qwen. Source froze before fresh transfer questions were written. All 186 formal replies were fully inspected, including failures, code and extra assertions.

Fresh authored transfer	Previous conversation R2	Conversation R3 / 0.3.9 source
Useful answers	10/29	22/29
Boundary handling (clarification/refusal included)	7/12	12/12
Complete conversations containing positive work	2/14	9/14

The candidate passed 332 inherited tests, 18 subtests and 124 focused checks. Repeated runs are not additional unique abilities. These are same-author, adaptive/unblinded development-family results, not standard benchmark accuracy, an independent Qwen win, or complete runtime/security certification. Evidence and all reviewed transfer failures are under benchmarks/conversation_039_r3/. Release packaging/integration checks are reported separately in RELEASE_0_3_9.md and the release manifest.

Remaining limits

Some selection wording repeats a stale choice. Some dictionary guides lose their topic on follow-up; richer examples and multi-definition rechecks can still fail. Source retrieval is lexical and incomplete. A link does not make an answer verified. Supported forgetting is not whole-transcript erasure or a general privacy guarantee. Finite grammar/VM coverage does not establish arbitrary English/Python understanding.

Earlier broad comparisons did not establish general superiority over Qwen; an earlier execution-policy deviation and session-contamination findings remain disclosed in their original evidence. State tracking development is not a clean StateBench victory. No new external benchmark score or live-web accuracy score is claimed for this release. AI-assisted development and unblinded review apply.

Download counter

The live badge above displays Hugging Face's counted downloads for the last 30 days. It is not a unique-user count or a count of every completed installation.

For a full repository download, use the official Hub CLI:

hf download vynly/CAEL-0.3.9 --local-dir CAEL-0.3.9-download

This includes params.json, a real CAEL download configuration containing the source/archive versions, pinned artifact URLs, sizes and SHA256 hashes. The Hub's standard download-counting rules include this filename. Extract the source ZIP and follow the installation instructions above; this remains a custom Python runtime, not a Transformers checkpoint.

Direct downloads of only .zip or .cael files, including older installer flows, are not comprehensively covered by the default Hub counter. Past archive requests are not reconstructed, and the displayed number can remain zero until Hugging Face refreshes its statistics. No extra tracking calls, telemetry, fabricated counts or simulated downloads were added.