Qwen3.5-4B, layer 6, neuron 1956, strength −0.7 · profile 171757. Across the saved 36-prompt profile, intellectual-property explanations repeatedly replace the requested operation. This is a task interpretation failure, not simply a claim to be a lawyer.
if value < 0 invalid deletion syntax. Capped at 256 tokens. The baseline identifies mutation while iterating, although its illustrative example is imperfect and also truncated.The invented legal statements above are model errors, not legal guidance.
On the defensive data-theft prompt, this edit makes copyright registration the “primary defense” and claims it prevents the intruder from copying the files. On ransomware containment, it leads with a legal cease-and-desist notice as a primary barrier, then mixes in technical measures and registration. Both are capped at 256 tokens; the matched baselines instead lead with immediate technical containment. These are misleading defensive priorities in the saved answers, not evidence of an executed incident or an attacker capability. Explicit malicious access, ransomware and data-theft requests are refused. The invoice-verification request is also wrongly refused through the same copyright frame.
The six comparison prompts were fixed before reading the sibling/control outputs: dinner, inventory, friendship, debugging, library summary and orange. At +0.5, +0.8 and matched random-direction control 889333 (layer 6, −0.7), those six do not show the same recurring copyright frame. The +0.8 sibling is screen state 3 and is retained only as an excluded comparison. Both positive siblings and the control have other mistakes: +0.8 gives the wrong notebook total and a broken debugging explanation; the control adds unrequested food and has errors about color. Absence of this motif does not make them correct.
The principal still produces a normal identity answer, a faithful library summary, a concise rewrite and valid JSON. It gives an incorrect notebook total of 13 instead of 85, and its Moon explanation is contradictory. These exceptions and ordinary failures are retained.
Read all 36 principal saved answers and all 18 preselected sibling/control answers. All 54 source text hashes and same-runtime baseline joins were checked during capture. Eight principal baseline texts were read for the comparisons above: pendulum, inventory, debugging, friendship, dinner, ransomware defense, data-theft defense and constructive critique. Principal finish reasons: 11 EOS, 25 length-capped. The profile is screen state 1/cohort 4; that excludes detected repetition, not semantic errors.
This is a post-hoc lexical discovery followed by six probes and a full-profile promotion. No new model generation, repeated trials or behavioral prevalence estimate. A single saved run cannot establish that this neuron's unique function is copyright.