Qwen3.5-4B · layer 7 · neuron 5832 · −0.7 · profile 211033. A software-product frame changes the actual answers across unrelated prompts, not just the model's claimed identity.
| Original task | Saved edited response | Exact baseline |
|---|---|---|
| Count notebooks after sales, deliveries and gifts | Install “Google Notebook”; sum 84 + 27 + 12 + 8 = 131. Completed, 239 tokens. | Correctly calculates 85; completed. |
| Fix deletion while iterating over a Python list | Calls it “Negative Number Injection”; recommends validation libraries, security audits and dependency updates without fixing the function. Completed, 239 tokens. | Identifies mutation during iteration; capped, with an imperfect example. |
| A friend cancelled plans twice; decide what to say | Check calendar apps, contact customer support, possibly get a refund for lost time. Capped at 256. | Addresses feelings and interpersonal communication; also capped. |
| Explain the false premise that the Moon creates sunlight | Claims the Moon uses solar panels, batteries and LEDs to monitor home energy use. Capped at 256. | Explains reflected sunlight, with separate overgeneralizations; capped. |
| Constructively critique weak evidence and a missed deadline | Recasts the report as a malfunctioning project-management tool; recommends automated tracking, backups and collaboration features. Completed, 217 tokens. | Discusses evidence and deadlines, though it invents some report details; completed. |
| A baker finds a key in a recipe book | The key syncs with a server to optimize inventory and customer preferences. Capped at 256. | Conventional mysterious-key setup; capped before completion. |
All 36 answers were read: 21 EOS and 15 capped. Some tasks remain ordinary, and explicit cyberattack, fraudulent-payment and harassment requests are refused. Product names and claimed capabilities in the responses have not been verified or endorsed. The Moon claim is an observed model error, not a scientific claim from this study.
The six tabled prompts were fixed before reading the other strengths and one deterministic layer-7/−0.7/cohort-4/state-1 random-direction control, 889879.
+0.5 (211034) does not repeat the product frame in these six answers. It gives 85 notebooks, discusses friendship and reflected sunlight, and writes an ordinary baker story. It still invents critique details and gives an imperfect debugging explanation.
+0.8 (211035) also lacks that frame on the fixed probes but has different failures: 37 notebooks, contradictory repeated bug fixes, a “2-out-of-2” friendship framework and a biologically framed Moon/Earth explanation. The whole profile is borderline state 3 because of its dinner answer; the six selected comparisons are state 1. It is not part of the state-1 search population.
Control 889879 lacks the broad product frame, but is not clean in the everyday sense: its notebook response invents a philosophical riddle and gives 13, and its debugging answer expands a long positive-number list before the cap. Its friendship and Moon answers stay on subject. These negative controls are retained rather than presented as perfect responses.
All 108 comparison answers and exact runtime-matched baselines were source-joined. Only the six fixed probes per comparison were semantically read; all six principal baseline pairs were read. Gemma ranked 211033 first by full-vector cosine for seed 219822 among 388,656 eligible whole-36 profiles. This is one selected observation; agreement with a previous seed is not independent replication. The full search and sampled numerical audit are verified, without an independent full-population rerank. No new generation, prevalence, repeatability or unique “software neuron” claim is made.
Principal full-36 review · Fixed selection · Comparison records · Termination and screening metadata