The Entitlement Cut: Reproducible Entitlement for Consequential Autonomous Systems, with the Autonomous Declaration Kernel as Derived Normal Form

Consequential autonomous systems can produce decision artifacts from which independent evaluators recompute recorded verdicts. Replay alone, however, establishes only that a result follows from its recorded inputs. It does not establish that those inputs—especially the governing regime, evidence boundary, authority lineage, and consequence rule—were already governing before the candidate action was selected. This paper introduces the Entitlement Cut, a necessary boundary for independently reproducible entitlement to consequential autonomous action. The central distinction is between entitlement as a historical property and reproducibility as an evidential property. An act is entitled when it is evaluated and authorized under a previously adopted entitlement constitution whose governing rule is protected against same-event retroactive amendment. Entitlement is independently reproducible when a separate conforming evaluator can establish that historical fact from the artifact under the declared replay equivalence. The main theorem proves, relative to a minimal bounded autonomous-consequence class, that any artifact supporting independently reproducible entitlement must identify a version-bound entitlement constitution and sufficient governance order to distinguish prospective adoption from a rule rewritten after the candidate became known. The candidate may legitimately influence which predeclared basis applies—for example through jurisdiction, hazard-class, or threshold routing—but it may not amend the rule selecting that basis for its own licensing event. The result operates over a control quotient rather than an implementation diagram: component count is irrelevant, and one process may preserve entitlement while a distributed system remains self-licensing if the candidate path controls the governing conditions. Temporal evidence preservation, basis-transition lineage, authority separation, typed failure localization, and the Autonomous Declaration Kernel are derived as manifestations or normal forms of the same boundary. A published finite-model audit exhaustively tests constant precommitted, candidate-sensitive precommitted, and candidate-responsive mutable selectors. Without governance-order lineage, three artifact classes mix entitled and gamed histories; with the Entitlement Cut bound, no mixed classes remain. The audit also refutes two predecessor formulations and provides a first model-relative analysis of the artifact field basis. The machine results are graded APPLIED; proof-assistant mechanization, field minimality, richer adversarial histories, transition composition, and full fifteen-role corpus integration remain open. The paper’s central compression is: Replay shows that the result follows from the selected basis. Constitutional replay shows that the basis follows from a previously governing rule. Non-retroactivity shows that the candidate did not rewrite that rule for itself.