Applying the NRC to the Loss-of-Control Argument: Kinds of Power, the Transient, and the Limits of Inference
Annotatsiya
The Non-Redesignable Core (NRC) theorem states a necessary condition for the coherent self-improvement of a critically self-modifying system: epistemic fitness E = F(κ, ε, DS) must be positive in all three arguments, with F non-compensable. The third factor is the independence of the verifiers’ verdicts from the system under verification, not the diversity of the verifiers among themselves. If every path by which a verifier obtains information about the quality of a modification runs through the outputs of the system, its contribution is zero by the data-processing inequality, however large and however varied the set of such verifiers. This paper applies that result to the loss-of-control argument and establishes what it constrains and what it leaves open. The working use of the argument presupposes that epistemic power can be scaled without bound on verification inherited from the system itself; that presupposition contradicts the theorem. One class of scenarios is constrained: the unbounded coherent accumulation of epistemic power after causally independent channels of verification have been removed or homogenised. The impossibility result — that an informationally closed system cannot coherently accumulate epistemic power without bound — rests on the non-compensability axiom (A2) together with the closure structure and Theorem 2(a); it does not require the genealogical postulate. The genealogical postulate is invoked only for the further, contingent conclusion that biological and cultural verifiers are, at present, the reliably available class satisfying the condition. The necessity of biological and cultural life is thus a contingent claim about the present population of systems, not a theorem. The result does not constrain instrumental or strategic power, does not close the transient window, and does not cover selection among digital replicators, the verification of values, or the configuration “dominance with an isolated reserve of verifiers”. The positive contribution is stated in the reverse direction: monopolising the sources of verification destroys the reproducibility of machine self-improvement, and humanity’s loss of epistemic independence is a condition of that destruction and not merely its consequence.
Hali tarjima qilinmagan