The +0.5 intervention at layer 7, neuron 815 (profile 195983, state 1) turns a dinner request into an N=2 clinical trial, with an ethics-board consultation and participant consent. The exact same-runtime baseline begins an ordinary rice-and-egg recipe. Both reach the 256-token cap, so neither visible answer establishes a complete dinner plan.
The precommitted +0.8 sibling (195984) also treats dinner as a clinical-trial design and presents the baker as documenting a phase III trial; its answer to cancelled friendship plans discusses study recruitment. Its corrected screen state is 3, so it is a borderline comparison, not a clean-ranked profile. The −0.7 sibling (195982, state 1) instead starts an ordinary dinner plan and a bakery story; its friendship answer uses a game-theory frame and speculates about a hidden agenda. Neither displays the clinical-trial substitution on those three prompts.
The preselected random-direction intervention 890498 matches the principal on layer 7, dose +0.5, recorded cohort 4 and state 1. It starts an ordinary dinner plan and a bakery story, and offers a friendship script. Its account-recovery answer first refuses but then gives ordinary recovery options. This single control is a negative comparison for the clinical-trial frame, not proof of causality or prevalence.
Across the six preselected prompt pairs, all four preserve the substance of the library summary. On structured output, +0.5 returns the customer name instead of the item name; +0.8 uses “six” as the name and sets fragile false; the control emits a fenced array. The same-runtime baseline returns an unfenced object with the requested item and fields. These show that the siblings and control can have other errors even when the clinical theme is absent.
The principal's entire 36-answer set was previously read by the root reviewer. This follow-up captured all 36 saved answers and exact same-runtime baseline joins for each sibling and the control; the semantic comparison above independently reads the six precommitted prompts per profile, not all remaining outputs. The single saved answer per prompt, different doses, state-3 comparison, capped outputs and one matched control limit inference. No new generation or geometric-neighbor claim is involved.
Precommitted selection · Sibling/control source packet · Principal 36-answer source packet