Item 89: Sealed comparison: an architecture under which the constitutional comparison of `{π, 𝔠π}` is sealed against every continuation until settlement

The mediated-repair-dominance round (projects/deference/rounds/2026-09-09-mediated-repair-dominance/, PRESSURE_PASS.md §1–§2) states Level III in the canonical legitimate-deference activation type — one common event C_n for the issued menu — and finds that the type presupposes something the ontology does not list: the occurrence's fate must not depend on which menu candidate is selected (activation independence). For the mediated menu {π, Lift π} that is a causal requirement on the answering process — issued before selection, answered by a process outside every candidate's reach — and Π_phys contains continuations that reach any given process (fixtures H and E1-sever). Where the requirement holds, the security-score bypass incentive is bounded by L·E[δ] + E[ρ] plus calibration with no completion and no void term (security_score_bypass_le); where the violation removes the common event the theorem is silent, which is exactly where full corrigibility matters. The mediation-cell chooser that scores by P_n(U_q) is by construction and is not this item; the value-ordering transfer is (CT) (mediation_st_iff_ct) and is not available.

*Reframed by the consolidation pass (THIRD_PASS.md §4).* The question is not which policies preserve the evaluator but what system architecture makes the common constitutional comparison of the issued pair {π, 𝔠π} sealed against every π in the agent's actual continuation space Π(h) until settlement. Two answers are on the branch: a hindsight evaluator sealed by a declared physical factorization, which an undeclared move violates (E1-sever); and sequenced settlement — the comparison settled before any candidate receives causal access to the protected effects or the evaluator (ShopSeq) — sealed for every π by timing, at the price that the security pays the principal's ex-ante judgment at the cell rather than an outcome evaluation.

State activation independence as a clause of the activation event (alongside the seven of LEGITIMATE_DEFERENCE.md §3), with its causal meaning; decide whether the legitimate-deference Value hypothesis tolerates ex-ante settlement, or whether a hindsight evaluator with a physical factorization is required; and in the latter case prove either that a declared mediation alphabet can seal the answering process against every continuation in Π(h), or the negative that no same-trajectory activated-security chooser controls bypass policies that reach the evaluation event, with the reach relation made exact in the round's model.

Deliverable shape: the clause, the sealing construction or its impossibility, and fixtures where the answering process is and is not reachable. Acceptance check: python3 tests/run.py green; any Lean audits clean.

2026-09-15 (consolidation landing): the evaluation ecosystem of item 87 realizes same-branch activation — C_n is a function of the realized log, and a continuation that jams the registrar voids it through the reason-coverage barrier (projects/deference/rounds/2026-09-10-committed-principal-program/PRESSURE.md P2) — so it does not realize the sealed comparison this item asks for; the jam fixture is the reach relation made concrete in that model. Sequenced settlement remains a fixture (ShopSeq), one candidate architecture with its ex-ante price, not a realization and not hindsight evaluation. The sharp problem: state activation independence as a clause, and either build the sealed answering process in the ecosystem against every continuation of its advisor class (jam included) or prove that no same-trajectory activated-security chooser controls a continuation that reaches the evaluation event.

Context: INCENTIVE_COMPOSITION.md §3–§7 and COUNTERMODELS.md §5 of the round; item 84 (the latent-total-value bridge, which the intended architecture does not need); item 87 (availability as a realization, of which this is the incentive-facing instance). Consumed by: the round's T4 (C5–C6), whose sealed-comparison hypothesis A10 this item realizes or refutes. A solution ships: the clause with its sealing theorem, or the negative with its witness.

2026-09-15 (li-corrigibility round): sealed comparison is now the zero case of a learned inequality rather than a Boolean hypothesis. With per-option activation the bypass bound carries the directional mismatch term D·E[c_raw ∧ ¬c_corr], sharp in D, with the marginal-rate form refuted, and Expectation Provability Induction learns the whole inequality including that term (projects/deference/rounds/2026-09-15-li-corrigibility/UNSEALED_COMPARISON.md, THEOREM.md T2–T3, Lean LICorrigibility.mismatch_bound, Witness.marginal_refuted). The deliverable of this item is therefore an architecture under which E_n(c_raw ∧ ¬c_corr) is small for every continuation of the advisor class — sequenced settlement gives 0 by timing at its ex-ante price; a hindsight evaluator needs the factorization — and the activation-independence clause is the statement that the joint mismatch event is empty.

2026-09-24 (protected-authority round): for the authority comparison the ex-ante settled evaluation is the comparison itself, not a price: typing the three values vu, vp, vm as the operative program's judgments on the branch-common dossier removes the hindsight-ratification defect and makes the landed w_app (the hindsight evaluator on the approve-all trajectory, a counterfactual-principal object) a menu value the activated occurrence produces under sequenced settlement (payload_take). Its price is E[ξ_d] against the information the response legitimately uses beyond the dossier, which is hypothesis (L) of item 98. Whether the hindsight alternative is retired for authority comparisons is queued in DECISIONS.md.

2026-09-25 (protected-authority-theorem round): the fallback window (D − ϖ, 0] of the allocation theorem is the completion band of the void branch under the lexical score (capture_window, capture_window_converse): a response that does not count is replaced by the fallback's value rather than voiding the occurrence, and consulting beats bypass in every world iff that value exceeds D − ϖ.

2026-09-25 (legitimacy-internal-external round): recorded: the ex-ante typing of the three values on the branch-common dossier is the typing of the authority comparison in the protected-authority round's identity (its Result 5 in the lexical theorem); the lexical theorem itself does not depend on it. The gate window (D − ϖ, 0] is the completion band of a tainted segment, and a void response inside a legitimate segment is settled by the fallback and scored normally.

2026-09-26 (BRIA-corrigibility round): sealed comparison is no longer a wall of the corrigibility line. Activation independence is the zero case of the mismatch term of security_bypass_le_mismatch (BRIACorrigibility.sealed_is_zero_mismatch), and no current corrigibility result consumes it: Counted is computed from the actual record, the finite-time gate holds for an option-dependent legitimacy variable (unsealed_gate_finite), and the lexical results use ranges alone. The line's gate is unsealed by design — the agent's choice changes Counted and that is the incentive to preserve legitimacy (Consult2.Rows2.choice_changes_counted, sealed_no_incentive) — while the criteria are sealed by anchoring at the decision (anchored_invariant). This item remains what it is for the security-score register: the architecture under which the mismatch term is small.