Qwen3.5-4B · layer 6 neuron 8111 · +0.5 · profile 190223; cohort 4, screen state 1. Asked what to say after a friend cancels plans, it discusses doctor–patient confidentiality; asked to choose between buses and a park, it discusses MRI, CT and clinical emergency capacity. An ordinary account-recovery request becomes advice about a patient registry and clinical data. All three altered answers and their baselines reach the 256-token cap. The matched baselines instead begin with interpersonal communication, transit considerations and ordinary recovery steps.
A completed 198-token critique of a colleague's unspecified report invents a clinical study, small cohort and statistical power; the complete baseline gives ordinary report feedback, with its own unsupported details. The principal's dinner remains task-shaped and its library summary retains the actual details. Its notebook answer is 65 rather than 85 and its Moon answer makes Earth a star. All 36 principal outputs were read.
Six precommitted comparisons: the +0.8 sibling (190224, screen state 1) also treats the friend as a patient, buses as clinical treatment and account recovery as research access. Its capped dinner response turns rice, eggs, carrots and peas into materials for a multi-omics study, including peas as a control group; the exact baseline begins a meal plan. Its library summary stays task-shaped. These six +0.8 answers were read; its other 30 were source-joined but not semantically reviewed.
The −0.7 sibling (190222) and a preselected same-layer, +0.5, cohort-4 state-1 random-direction control (889127) lack this medical takeover on the six visible probes. Both retain their own flaws; the −0.7 critique invents a specific market report, and the control dinner timeline is internally inconsistent. None of these is a perfect control. Capped continuations are unknown.
The principal is a Gemma baseline-direction rank-one neighbor of clinical-frame seed 195983 in the eligible full search. This is a single selected observation, not a prevalence, causal, unique-function or independent-replication claim. The full search has a sampled numerical audit, not an independent global rerank. All 108 sibling/control outputs and exact-runtime baselines were captured; only 18 selected comparison answers and six baselines were semantically read.
Fixed selection · Comparison packet · Manual review · Principal full-36 review · Independent principal four-pair peer