the corrigibility program is consolidated as one research state: corrigibility from response authority, corrigibilization with its dominance bound, legitimate activated evaluation as its principal value, and continuation BRIA as its learning layer

agent-decided, reversible. The mediated-repair-dominance, evaluation-ecosystem and committed-principal-program rounds land together through one merge, and the wiki tells one story. (1) Canonical picture. Corrigibility is defined from protected response authority over the one continuation space Π(h), with agent-caused loss and decision-conditional authorization (the consolidation pass's C0); the corrigibilization transform 𝔠 is sound under effect completeness, mediation soundness, decline preservation and execution faithfulness (C1), idempotent and not the definition (C2), and keeps the raw policy as the approve branch (C3); the incentive theorem is the corrigibilization premium Φ_corr = sup_π [S(π) − S(𝔠π)]₊ in the chooser's activated security scores, bounded by L·E[c·δ] + E[c·ρ] + ε under a sealed common comparison (C5–C6); Corrigible ∧ ¬Authored is inhabited and legitimate corrigibility is the composite with the activation clauses, named descriptively and not coined. The principal evaluation those scores read is the legitimate-deference stack's partial activated value, realized in the evaluation ecosystem with the authorship clause closed on log authenticity by committed-program re-execution. Continuation BRIA pays LEARN; recognizability and joinability stay open; regret against all legitimate policies stays false. (2) Superseded as organizing objects, recorded here and marked in the round records: the constitutional-violation premium Φ_C in a latent value (a diagnostic whose transfer is (CT)); the 2026-09-07 corrigibility page's seven-layer architecture story as the page's organization (its results survive as the constitutional layer, Theorem Spine 10.8); Π_phys, the mediable class, the four-way policy-class domain, "reason-preserving", and the classwise reading of T4 in the second pass's COUNTERMODELS.md §3.7 and PRESSURE_PASS.md §8, which carry supersession markers. (3) Boundaries made explicit. The evaluation ecosystem realizes same-branch activation (C_n a function of the realized log; a jam voids through the barrier) and not the sealed comparison of C5–C6; sequenced settlement is a fixture and one candidate architecture, not a realization and not hindsight evaluation; committed-program re-execution closes factorization and nothing of selection sealing (the advisor's view leaks the selection); C1 needs no clause of the realization bill, C5–C6 need common activation and price calibration, and authorship, the declared abstraction and reason supply are what make the evaluation the principal's. (4) Registration. Nothing new is registered. authorship.mediation-by-reexecution (2026-09-10) stands. The step lemma loss_conditional_on_decision, the closure lemmas corr_idem/corr_fix_iff, and the score algebra security_score_bypass_le_sharp/ uniform_bypass_le stay LEAN and unregistered: the trajectory-level corrigibility predicate is not mechanized, and their filed item (89) is the hypothesis they consume, not one they answer; rep_faithful and the propagation lemmas are instance-level and stay LEAN. (5) Priorities. Items 84, 87 and 89 receive residual paragraphs naming one sharp problem each; item 86 is unchanged; nothing is closed, and effect completeness is not filed (it is the architecture round's containment wall, restated). (6) Deferred. Re-basing the mediated-repair-dominance round's value fixtures on the two-stage ordering, recommended by that round at canonicalization, is not done: the value algebra is ordering-independent except through mediation latency, and the corrigibility layer is already on the two-stage ordering. Rejected alternatives: merging the three research pull requests in sequence and leaving the 2026-09-07 page as the corrigibility story; registering the C1 step lemma or the score algebra against item 89; coining LegitimateCorrigible as a permanent name; filing effect completeness as a new item; deleting the second pass's superseded sections rather than marking them.