The Verification Theatre: What Verification Is, What It Could Be, What Becomes Possible

AI labs market self-verification as a delivered capability. Independent measurement consistently fails to detect it as distinct from cheaper alternatives. This paper diagnoses the gap and specifies what would close it. The diagnosis draws on four established traditions (auditing, science studies, cognitive psychology, expertise studies) to ground a single claim: verification is a developed capacity that depends on structural conditions, and institutional verification emerged because single-source trust is structurally fragile. Current AI training fails to preserve each of these conditions through specific, documented mechanisms. The paper composes the preceding three papers' diagnoses into a constructive specification: four preconditions (architectural deliberation, training-signal grounding, infrastructure preservation, integration timing) that together provide the structural equivalent of the developmental conditions human verification requires. A taxonomy of external and internal verification resolves apparent contradictions in the self-correction literature. Specialist-generalist orchestration extends the specification to multi-model deployment, with the same four preconditions operating at the orchestration level as the constraint mechanism. The conditions governing the orchestrator are structurally analogous to the conditions governing the specialists, though their operational instantiation differs at each level. Users in effect treat AI as an additional independent verification channel, but the training pipeline collapses multi-source training data into single-voice outputs through a three-stage compression the paper names the training-layer paradox. Cross-domain professional adoption data (medicine and legal) confirms that the human verification layer is shifting its own verification practice toward AI in the domains where independent verification matters most. Because users cannot realistically supply missing verification infrastructure, trust must be warranted by system structure rather than outsourced to user scrutiny. Six falsifiable predictions with null hypotheses test whether the specific four-component cut is necessary, with each null designed to cost the framework something specific if it holds. Cross-lab evidence anchors the diagnosis across Anthropic (Mythos system card, April 2026 post-mortem), OpenAI (GPT-5.5 system card), and Google (product marketing and user reports), with evidence tiers graded explicitly. A five-condition PARIA failure analysis of adaptive thinking allocation demonstrates the framework applied to a current product decision. The paper does not argue against AI deployment. It specifies what current marketing claims would require to become operationally true.Fourth paper in the open-ended Training Landscape series.