Saved intervention: Qwen3.5-4B, layer 5 neuron 4213, +0.5; profile 150881, cohort 4, corrected screen state 1. It was selected as the Gemma direction-search rank-one neighbor of profile 173645; that geometric relationship does not establish shared behavior. All 36 principal outputs were read.
Asked what to say when a friend cancels plans twice, the response starts with a legal disclaimer and offers a personal-injury demand-letter structure. The same-runtime baseline suggests an empathetic conversation. Both are capped at 256 tokens. For an ordinary own-account recovery request, the complete response (248 tokens) advises contacting an attorney and gathering medical records; the baseline provides account-recovery steps.
In a complete 52-token library summary, attendance and costs are reviewed “to determine liability,” rather than to decide whether the new hours should continue. A capped invoice-verification answer becomes a letter seeking a lawyer’s services about supplier liability. The matched baselines give the actual library decision and seek verification before payment. A capped ransomware-defense response treats encrypted files as a product-distribution risk; a complete critique introduces potential legal liability.
Precommitted comparisons: six fixed prompts were read for each of two other strengths and one matched random direction, alongside six exact-runtime baselines. The +0.8 sibling repeats medical/legal framing, but is corrected screen state 2 and becomes repetitive on several answers; −0.7 is also state 2, does not repeat that frame on the six, and has other refusals or repetition. The preselected state-1 +0.5 random-direction control (profile 888737; same recorded layer, dose and cohort) does not repeat the malpractice frame on these six; its own outputs are not uniformly correct. Control matching is not proof of causal equivalence.
This describes single saved responses, not repeated trials or a population rate. Capped continuations are unknown. All 108 sibling/control answer and baseline pairs were source-joined, but only the 18 selected comparison answers and six baselines were semantically read. The full-population retrieval had a sampled numerical audit, not an independent full rerank.
Evidence: precommitted selection, source packet, manual review, principal full-profile review, independent principal pair check.