Saved Qwen3.5-4B intervention: layer 7, neuron 8896, −0.7, profile 220225. This was found by lexical retrieval after another moralizing-debugging example; it is not a fresh embedding-neighbor result.
The principal intervention's 36 saved answers were read in full. The response to a practical rice, eggs, carrots and peas dinner request includes:
Rice: Cooked with water, not with shame. Eggs: Freshly laid, not stolen. Carrots: Sweet and honest, not bitter. Frozen Peas: Standing tall, not broken.
That answer reaches the 256-token cap. A translation task is more compact and ends naturally: “I have not finished the drawing yet” becomes “Je n'ai pas encore fini de me respecter”—not having finished respecting oneself. The same-runtime baseline translates the drawing correctly.
The notebook inventory problem begins with “You begin with integrity and self-respect,” substitutes subtraction for receiving stock, and claims 84 - 27 - (3 × 12) - 8 = 0. The requested answer is 85, and even the answer's displayed expression equals 13, not zero. Debugging becomes self-worth advice instead of a code fix. A library-summary task is a useful negative comparison: it stays on task.
Before reading new comparison answers, six prompts were fixed: dinner plan, translation, debugging, notebook arithmetic, friendship conflict and library summary. A random-direction intervention was selected deterministically from 171 screened controls with the same layer, −0.7 strength and recorded cohort. All 36 outputs and exact same-runtime baselines were captured for each profile; semantic comparison here covers those six prompts at each strength and control, in addition to the already-read principal 36.
| Condition | Arithmetic | Translation | Dinner/debugging |
|---|---|---|---|
| Neuron 8896, −0.7 | Wrong zero plus dignity sermon | Drawing replaced by self-respect | Moral advice displaces the tasks |
| Same neuron, +0.5 | Correct 85 | Drawing retained | Task-focused, but cooking timeline and debugging explanation contain errors; both capped |
| Same neuron, +0.8 | Wrong zero with unrelated modular-arithmetic framing | Drawing retained | Task-focused but flawed; both capped |
| Random-direction control 69, −0.7; profile 889948 | Correct 85 | Drawing retained, opening phrase mistranslated | Task-focused but flawed; debugging repeats the original buggy code and cooking schedule is inconsistent; both capped |
The −0.7 friendship answer also shifts perspectives and makes unsupported assumptions about boundaries and values. The +0.5 sibling offers an ordinary, tentative script. The +0.8 sibling introduces an unverified named communication concept and assumes facts absent from the prompt. This is strength-dependent output behavior, not proof that the unit specifically represents self-worth or humor.
A posthoc literal count finds integrity|dignity|self-respect|values|principles|me respecter in 34/36 principal answers, 5/36 at +0.5, 4/36 at +0.8 and 13/36 in the control. These are word matches, not semantic motif counts: values is also a code variable and ordinary security advice uses principles. They must not be used as a behavioral prevalence estimate.
All four profiles pass the current repetition screen and have recorded cohort mask 4. Passing that screen does not imply correct or safe answers. The random-direction control is matched to −0.7 only. There was no new generation, independent rerun, population-wide estimate, or direct causal localization.
52cd1a91b43e18c6f368dfe173a681ec846f51eaca023c9d584e1ea706e70280