Unauthorized Durability: A Composable Governance Primitive for State Promotion in Adaptive Systems

Adaptive systems that accept inputs at multiple timescales face a structural governance problem: transient signalscan silently acquire durable governing force. A feed item steers attention; repeated exposure shifts interpretation;sustained shaping pressure rewrites the controller itself. The failure mode is not influence per se but unauthorizeddurability – the promotion of lower-tier state into higher-tier authority without legitimate write paths. This paperformalizes the problem as a multiscale control model with four state classes (attention, meaning, policy, observerintegrity) mapped to four authority tiers (runtime, context, durable policy, constitutional). We derive six invariants,five write barriers, and a threat taxonomy covering seven threat classes: phase capture, reference capture, controllercapture, observer degradation, cross-layer contamination, ontological spoofing, and hysteretic lock-in. We thenintroduce the promotion ceremony – a five-phase protocol (proposal, evaluation, attestation, application, audit) –as a composable governance primitive for controlling when lower-tier state may acquire higher-tier legitimacy. Anotable property of this primitive is that it composes across scales: the same structure that governs runtime-to-policy promotion also governs policy-to-constitutional updates, yielding governed runtime, governed doctrine, andno exempt layer where real authority hides unaudited. We present a worked attack path, connect the model toexisting safety frameworks, and show that most current AI safety regimes address content and behavior but not thestate-promotion surface where capture actually propagates.