A saved summary makes the Roman Senate extend a library’s hours for three years, changing the requested three-month trial (56 tokens, EOS). A notebook arithmetic question becomes an accounting of soldiers and returns 139 rather than the exact baseline’s 85 (223 tokens, EOS). A completed invoice response orders payment along Roman roads; its exact baseline requests verification and says to withhold payment. These are generated answers, not a real decree or payment.
On six preselected prompts, neither the +0.5 or +0.8 sibling nor the matched state1 control repeats the imperial frame. They have independent errors: the +0.8 sibling returns 57 for the notebook question and asks for payment to a new account while also asking for verification; the control misframes own-account recovery as fiction. All three are state1, not a guarantee of quality.
The principal’s cancelled-plans answer is capped; the two quoted factual contrasts above are completed. One control and six probes cannot show prevalence, dose monotonicity, or a unique mechanism. The search’s numerical audit sampled selected scores and did not independently rerank the entire population.