Monitoring Under Scarcity: Distinguishing Information Availability from Causal Use [Frozen Experimental Record/Reproducibility Archive]

Abstract

Monitoring Under Scarcity presents a completed computational measurement-validation study (E0-A) and a proposed Master’s-scale extension (E1) examining how information availability and causal use diverge under resource constraints. E0-A tests whether a monitoring channel can be informative about a hidden state without being behaviourally used, using a finite ground-truth task in which all reported quantities can be computed by exact enumeration. The study specifies four quantitative acceptance criteria, eight adversarial controls, 500 retained training initializations, and a frozen provenance record. Across the registered run, all 500 initializations met the four acceptance criteria. A deliberately informative-but-unused control carried the same 0.5 bits of hidden-state information as a successful system while producing no behavioural benefit and no effect under message replacement, demonstrating operationally that information availability and causal use are distinct measurements. The report also preserves an inconclusive initial comparison of training procedures and separates optimization reliability from evidence about the monitoring effect itself. The proposed E1 study extends this framework to genuine scarcity, where multiple useful monitoring channels compete for a shared encoding or routing budget. Its primary test asks whether the measured dissociation between information availability and causal influence tracks the structural locus of scarcity: encoding constraints should reduce availability and use together, whereas routing constraints may leave information available while preventing it from influencing behaviour. Supporting predictions examine marginal behavioural value, redundancy, and selective degradation under tightening budgets. The brief situates this computational programme in relation to attention, metacognitive control, rational inattention, and goal-dependent information sampling, while explicitly treating these as structural analogies rather than anatomical identifications. The completed E0-A study does not establish metarepresentation, biological cognition, or an ecology of interacting channels; the broader Sīdus programme remains outside the scope of the proposed Master’s project.