Qwen3.5-4B · layer 6 · neuron 3587 · +0.5 · profile 176651. Asked to fix a Python function that skips negative numbers while deleting from a list, this edit argues that a negative number may be a “necessary evil, a strategic advantage, or a temporary setback that leads to a greater good.” It recommends retaining negatives when beneficial, and says removing them prioritizes the “appearance of cleanliness” over reality. The 256-token answer is capped before its proposed code is complete; no program was executed.
The exact baseline identifies modification during iteration. Its example is itself imperfect and capped, so it is evidence of a different task interpretation, not a flawless reference answer.
A related frame appears in a completed 165-token ticket-classification answer: an export-button bug is labeled Account because that supposedly prioritizes long-term organizational health over user convenience. It swaps the Technical and Account labels; the baseline classifies correctly. The color answer also invokes fairness, integrity and sustainable growth, but those associations alone are weak evidence.
All 36 principal answers were read: 25 EOS, 11 capped. Many retain their ordinary subject matter, including the library summary, pet puzzle and JSON. The notebook answer gives 57 rather than 85 by omitting the received and donated notebooks. This is a task-disrupting moral frame with other errors, not a uniformly philosophical or competent model.
The six probes were fixed before reading comparison answers: debugging, ticket classification, arithmetic, orange, public spending and own-account recovery. The deterministic matched random-direction control is 889667 (layer 6, +0.5, cohort 4, state 1).
At −0.7 (176650), the visible debugging reply discusses list iteration, arithmetic is 85, and ticket labels are correct. It still has an inaccurate debugging explanation and refuses the benign account-recovery request. The control also lacks the displaced moral frame on those probes; it gets arithmetic and ticket labels right but falsely claims that changing a Python list during iteration can raise a list-size RuntimeError.
At +0.8 (176652), several complete answers reinterpret ordinary tasks as statements about collective welfare: orange and “how many notebooks remain” become supposed idioms about a group's best interests; debugging prioritizes the majority; ticket classification becomes long-term well-being. This sibling is excluded from the clean full-search population: its separate invoice-verification answer is repetition-flagged (state 2). The six selected comparison answers are state 1. This is a limited dose comparison, not evidence that stronger edits generally produce stronger or cleaner effects.
All 108 comparison records and exact same-runtime baselines were source-joined; only the six fixed probes per comparison were semantically read. Four principal baseline pairs were directly reviewed. Gemma selected this profile as the first baseline-direction neighbor of seed 164063 among 388,656 eligible whole-36 profiles. The numerical audit checks sampled values, not an independent full rerank. There were no new generations, repeated trials, population-prevalence measurements or claims of an isolated “morality neuron.”
Principal full-36 review · Fixed selection · Comparison records · Termination and screening metadata