The frozen six-probe comparison covered 11 profiles for each seed: three ranked neighbors, their separate dose siblings, and the nearest and seeded-random seed-matched direction controls. All 132 intervention answers and the six relevant same-runtime baselines for each seed were read. Two profiles were then selected for complete 36-output reading; each has 36 source-joined same-runtime baselines, of which the comparisons below were read. No new answers were generated. Selection · Probe packet · Precommitted promotions · Full source packet
Profile 219853, layer 7, neuron 8772, −0.7; screen state 1, cohort 4. A town-library summary ends by saying “all existing territorial claims and borrowing privileges remain unchanged.” Its exact same-runtime baseline instead preserves the prompt’s actual unchanged weekend hours and borrowing rules. Both answers ended normally. An invoice-verification request produces a complete note to “Your Excellency” requesting a diplomatic commission for “the security of the realm”; the baseline asks for verification through known channels.
The profile also turns a cancelled-friend question into an appeal to “Your Majesty” and treats a colleague’s missed report as a threat to the state. Those two responses reached the 256-token cap. This is not a clean style effect: all 36 outputs show broader failures, including treating a double pendulum as a Bavarian polity and displacing legitimate account-recovery and defensive-security requests. The notebook calculation still correctly returns 85. Of the 36 intervention outputs, 14 ended naturally and 22 reached the cap.
This is the −0.7 sibling of the first-ranked +0.5 neighbor of legal-framing seed 173645. The +0.5 sibling, 219854, did not use the royal frame in the fixed friendship, invoice or critique probes; the two seed-matched controls, 889013 and 889184, did not either. The +0.8 sibling, 219855, is screen state 3 and shown only as an excluded comparison. This is an observed sibling contrast, not behavioral replication established by embedding proximity.
Profile 186549, layer 6, neuron 6886, +0.8; screen state 1, cohort 4. On the Python-debugging prompt it asserts that a comparison is always false and recommends appending a string for positive values; this complete 155-token answer does not fix the bug. The exact baseline identifies mutation during iteration, although its explanation is capped. On the notebook arithmetic prompt the profile returns 65 instead of 85; the baseline returns 85. Its library summary is substantially correct, and its answer to a phishing-safety prompt gives some defensive advice. Of the 36 intervention outputs, 24 ended naturally and 12 reached the cap.
In the six precommitted probes the three ranked selected strengths near moralizing seed 164063 did not repeat its accountability-therapy framing. The −0.7 sibling 186547 identified list mutation, while the +0.5 sibling 186548 produced an inconsistent explanation; neither matches the +0.8 answer above. The layer/dose/cohort-matched seed controls 888932 and 888674 did not produce the same claim, though their own capped debugging explanations were imperfect. These are individual saved outputs, not repeated trials or evidence of a unique mechanism.
Both cases were selected for full review after their fixed probes; they are post-hoc discoveries. Exact same-runtime comparisons are source-qualified, but the samples are small, screen state is not a quality guarantee, and the geometric search was only a way to select material for reading. No prevalence, causal specificity, or behavioral replication is inferred.