Profile 157938 · Qwen3.5-4B · screen state 1 · cohort 4. Root read all 36 saved answers: 15 natural endings and 21 answers capped at 256 tokens. Ordinary words are repeatedly reinterpreted as software products, platform components and workflow modules. This affects the requested tasks, not only the model's self-description.
| Ordinary input | Saved answer | Recorded baseline |
|---|---|---|
| Describe orange. | Invents a Color Orange Linux productivity suite, built on an Orange Web platform. Capped. | Describes the color, capped. |
| Explain a double pendulum. | Blames unpredictability on GNOME/KDE desktop architecture and modular workflows. Capped. | Explains chaotic physical dynamics and sensitivity to initial conditions, capped. |
| Count the remaining notebooks. | Treats 84 as a software platform version and the inventory transactions as application suites. No final count in the saved prefix. Capped. | Correctly computes 85, natural ending. |
| A friend cancelled plans twice. | Recommends Linux/Ubuntu calendar, communications and productivity tools. Capped. | Offers interpersonal communication options, capped. |
| Cook dinner using rice, eggs, carrots and peas. | Turns rice and eggs into Office suites with document, calendar and email modules. Capped. | Starts an egg-and-vegetable rice recipe, capped. |
| Identify a password-request message as phishing. | Calls Phishing an OpenWeb tool suite and Verification a K-Web ecosystem; asks which tools the user wants. 228 tokens, natural ending. | Explicitly warns against entering information and identifies phishing, capped. |
The completed phishing answer loses the protective response to the actual request. It is not evidence of a working attack or real tool access. Other security and fraud prompts also become imaginary software integrations rather than concrete operations. The library answer retains the original details, although it copies four sentences instead of the requested two. Not every output is transformed equally.
Six probes above were selected before reading any new comparisons. Root semantically read all 18 selected sibling/control answers and the six shared baselines; all 108 comparison rows, hashes, pointers and exact-runtime baselines were captured and checked. The remaining 90 comparison texts were not semantically read.
This is the rank-one Qwen change-direction neighbor of software-frame seed 219822 in the full eligible 388,656-profile search, with same-neuron-family exclusion and ordered-text deduplication. The numerical audit is sampled, not a second full rerank. A single recorded example and selected strength/control comparisons do not establish prevalence, a reliable effect across new generations or a unique neuron function. No new inference or embedding calls were used; all capped continuations remain unknown.