Universal Turing Machines, Predictive Closure and Self-Revision: Implications for Reflective Consciousness

A universal Turing machine can revise the programmes through which it acts and learns, but sustaining a useful model of its own activity requires information to survive those revisions. I examine this requirement in an agent whose selector and updater can both be replaced during ordinary operation. Fresh task outcomes supply informative feedback, and subsequent predictions concern the continuation that the agent actually selects. An exact finite bound determines the best attainable probability of predicting that complete continuation from the information retained in model memory and executable code. For k independent fair outcome bits, constant code, a B-bit retained model and e observations erased at fixed locations, the optimal probability is 2^(-max{e, k−B}) under a complete-readout task. The general result accounts for observations that survive in revised programmes, showing why erasing an apparent model store need not remove their influence. A separate construction demonstrates how ordinary feedback improves later prediction and changes the effects of replacing both selection and updating procedures. These results complement an exact model-size theorem for all admitted revisions, with conditions for preserving the relevant organisation across implementations. They quantify information available to changing control procedures. Applying these constraints to reflective consciousness additionally requires that the predictive target concern the agent’s own organisation, rather than task outcomes alone, and that its physical implementation be established.