AI alignment is usually framed as retaining human control over increasingly capable systems. This paper studies the limiting case in which that control can no longer be assumed and asks what epistemic reasons remain available to a finite, world-directed agent under model uncertainty and nonstationarity. The core argument has three levels. First, when a model or a policy derived from it helps generate the evidence used to validate the model, apparent agreement can rise while the evidential weight of that agreement falls (Proposition 1). Second, a verification procedure expressed entirely within a model's current partition cannot certify exhaustive coverage of action-relevant distinctions outside that partition (Proposition 2). Third, under model uncertainty, substantially reversible policies retain update value unless they create a demonstrably larger catastrophic risk than irreversible alternatives (Proposition 3). Together these results motivate an architecture that separates prediction from validation, preserves independent observer classes and raw provenance, retains channels through which reality can contradict the controller, and treats irreversible elimination as requiring evidence not generated solely by the policy recommending it. These core claims neither require humans to be uniquely irreplaceable nor depend on the supplementary Compression-Fidelity Framework or Ownership Conjecture. The specifically human application is conditional on A7′: after the strongest genuinely independent artificial expansion available within a declared resource envelope, removing human witness channels must leave a predeclared material residual loss on at least one admissible decision problem. A separate moral ground applies only to agents assigning weight to consciousness or moral uncertainty. Empirically, Experiments 1-3 establish and refine the endogeneity mechanism in synthetic environments, while Experiment 4 confirms it with learned classifiers on real handwritten-digit measurements. Experiment 4F then tests observer-frame fidelity in twenty newly randomized twelve-witness portfolios after an internal freeze. All four 4F contrasts passed, but the equal mean beat every calibration-selected subset and stacking helped only after distribution shift. Experiment 4G supplied the programme's first public pre-outcome registration: Zenodo version DOI 10.5281/zenodo.22243509 was published before forty reserved portfolios were evaluated in diverse, homogeneous and exact-clone arms. All five substantive 4G contrasts passed their familywise 99% criteria. The diverse-minus-homogeneous structural gap was 1.6392 [1.5411, 1.7326], the corresponding three-bit penalty difference was.6120 [.5799,.6433], and the shift-specific stacking interaction was.1076 [.1024,.1127]. The exact-clone structural identity remains plumbing, not evidence. The adverse implication is equally important: substantial artificial frame diversity was easy to regenerate, and the controls locate the large penalty in the bundled diversity intervention without identifying which component causes it. These are bounded same-programme and restricted-class results, not independent replication, a frontier-language-model result, a universal rate-distortion theorem, or evidence that A7′ is true. The conclusion is therefore a burden of justification rather than a derivation of human supremacy: do not treat agreement from self-authored evidence as proof that independent, potentially irreversible sources of correction are replaceable. ( direct link )