sealed comparison is the zero case of a learned inequality

Agent-decided, reversible. The security-score bypass bound is stated with per-option activation and a directional mismatch term, E[U_raw] − E[U_corr] ≤ L·E[both·δ] + E[both·ρ] + D·E[c_raw ∧ ¬c_corr], and the common-activation ("sealed") case is E[c_raw ∧ ¬c_corr] = 0, not a hypothesis of the theorem. The rejected alternative — keeping sealed comparison as a Boolean system hypothesis (the 2026-09-09 A10) — was rejected because the bound with the term is sharp, its marginal-rate variant is false, and the Logical-Induction form learns the term along with the rest. Item 89's deliverable is restated accordingly (PRIORITIES.md).