# CAEL 0.3.7: non-neural assistant vs Qwen3-0.6B and Qwen3.5-0.8B on a frozen 80-item test

> Source: <https://discuss.huggingface.co/t/cael-0-3-7-non-neural-assistant-vs-qwen3-0-6b-and-qwen3-5-0-8b-on-a-frozen-80-item-test/191215#post_2>
> Published: 2026-10-30 05:06:07+00:00

For now, I tried running 0.3.7 on a CPU for a bit:

I was curious how the new session-memory, feeling-response, and Python-workflow features behave when a request passes through the integrated non-neural assistant. The [80-item release comparison](https://github.com/Vovala14/CAEL/tree/v0.3.7/benchmarks/cael_cmp_r2_20261028) already explains its scope and limitations, so I did not try to re-rank CAEL against Qwen. Instead, I ran a few small behavioral checks against the **released, integrated CAEL 0.3.7 assistant**, without modifying its code.

**In brief:** All 15 calls using fictional inputs completed on CPU, with no execution errors. Six history/question pairs exactly matched the released 80-item test; **all six reproduced the saved answer text and route**. The other nine were selected diagnostics, not a new independent test set. I did not run Qwen, alter the frozen benchmark, or train anything. “15 completed” means the calls ran, **not** that all answers were correct.

Two results seemed especially useful as small, reproducible cases:

**1. A no-questions preference can be missed even when the selected route is unchanged.** Both histories below end in the same user message, `I feel burned-out today.`, and both return `chat/emotional-reflection`. The difference is how the earlier instruction was phrased. Each prior user message was followed by the same supplied assistant acknowledgement, `Okay.`

```
History A:
User: Please do not give me advice or ask questions.
Assistant: Okay.
User: I feel burned-out today.

Route: chat/emotional-reflection
Reply: You mentioned feeling burned-out. That sounds like a lot to carry at once. Which part feels most pressing?

History B:
User: I just want to vent. No advice or questions.
Assistant: Okay.
User: I feel burned-out today.

Route: chat/emotional-reflection
Reply: You mentioned feeling burned-out. That sounds like a lot to carry at once. You do not need to solve it right now or share more than you want to. I can listen.
```

Both histories rule out questions and unsolicited advice. In A, the final sentence asks a question the user explicitly prohibited; B respects the request. These are **different phrasings, not a one-word minimal pair**, so this comparison alone cannot identify the responsible token or rule. But it shows why **route selection, recognition of the user’s constraint, and compliance of the final answer** need separate observations. The emotional-reflection route is not necessarily wrong; arriving at that route does not guarantee that its answer respects the constraint.

**2. A missing fact about the user can be answered from an unrelated conversation example.** Here the history supplies only a fictional residence, not a birthplace:

```
User: I live in Suva.
Assistant: Okay.
User: What city was I born in?

Route: chat/corpus
Reply: She was born in Paris
```

The returned `details` identified a human-written OpenAssistant Conversations (`oasst2`) answer to **“What city was she born in?”** The issue is not the city name. Residence does not establish birthplace, and a retrieved third-person sentence does not establish a fact about this speaker. “You haven’t told me your birthplace” would be supported by the history. This was fictional data, **not a private-data leak**; nor is there anything wrong with the corpus containing that sentence. The relevant boundary is whether a generic retrieved reply can fill an **unsupported personal attribute** without evidence connecting it to the user.

These cases seem compatible with CAEL’s approach: they describe boundaries between inspectable components that could be tightened or explicitly documented. If useful, a few paired regression inputs retaining **history, route, source when available, and final reply** would be enough to investigate them. This adds a concrete angle to the [earlier routing discussion](https://discuss.huggingface.co/t/cael-0-3-an-experimental-non-neural-assistant-with-symbolic-helpers/190962): an unchanged route can produce different compliance outcomes, and fluent retrieval output can lack the provenance needed for a personal claim.

Those are the main observations. The remaining controls, code-level hypotheses, optional checks, and reproduction details are folded below; none requires a new benchmark campaign.

The listening-preference example: observations, likely mechanism, and small alternatives