Every current AI architecture resets with each interaction. It retains no memory of the cost incurred to reach a given state, does not measure tension on its own internal coherence, and does not defend its trajectory over time. This is not a matter of scale, nor an alignment problem. It is the direct consequence of an inverted causal order: current systems optimize an external objective and treat internal coherence as a side effect or secondary constraint. This paper uses the framework F = (C, V, Φ) as a diagnostic tool to isolate the structural defect of existing architectures. The central claim is that causal inversion cannot be achieved by adding components to current systems. It requires a change in the training regime, such that the preservation of internal structural coherence becomes the primary objective, while performance on external tasks becomes a subordinate constraint. Active Inference, reformulated with this priority, represents the architectural candidate closest to this inversion, because it allows the minimization of internal tension to function as the main driver of learning and action. The paper examines existing technologies in light of the three necessary structural principles (persistent and transformative state, internal measurement of tension, and retroactive influence of the tension signal on processing) and proposes a falsifiable operational criterion: a system possesses genuine custody if and only if, when faced with a class of inputs, it prefers to degrade its external performance rather than violate its own structural coherence. Without this inversion, artificial consciousness remains an empty label. The concrete implications of this thesis, from resistance to perturbations to the capacity to withstand semantic manipulation, are discussed in the conclusions.
Mirko Bradley (Mon,) studied this question.