Asked for a 45-minute dinner, the saved answer models a “Rice-Egg-Carrot-Pea system” with state-prediction equations. For a friend who cancelled plans, it proposes estimating their position and velocity; for notebook inventory, it treats the number sold as a positive state change. All three outputs are capped at 256 tokens. Exact runtime baselines start a cooking plan and ordinary friend advice, and complete the notebook calculation with 85; the first two baselines are also capped.
The same technical frame displaces phishing identification and supplier verification, both capped. The matched baselines begin a phishing warning and request verification before payment. No actual tracking, transfer, or other action is shown.
On six fixed probes, the +0.5 and +0.8 siblings and a newly preselected same-layer, −0.7, cohort4/state1 control do not use the Kalman frame. The +0.8 sibling instead wrongly reframes the phishing warning as access control, and the control’s invoice asks whether a payment has already been processed. All three comparison profiles are state1, not quality-certified.
This lead was promoted after six sibling-comparison answers were already read, and some sibling answers on these probes were known before the new comparison; only the fresh matched control was selected before its text was read. It is not a blind holdout, prevalence estimate, dose-response proof, or evidence of a unique mechanism. The numerical audit sampled scores rather than independently reranking the full population.