Many systems contain predictive structure that is not equally accessible from all representations. Existing studies often conflate representation, information budget, and predictor flexibility when comparing how well a representation supports prediction, leaving it unclear which factor is actually responsible for an observed difference. ISRC-A proposes a framework -- Accessibility(Y|X) = F(C_theta, Z, f), not F(C_theta) alone -- that separates these factors explicitly, so a representation comparison can be attributed to the property that actually caused it. Two independent cases support it: in a physical system (Paper A, a 2D Ising trajectory ensemble), representation choice determined whether an existing predictive relationship was recoverable at all (scalar/discrete path, r=0.588 at best, no separation of a known special case; continuous compression, r=0.540, 74% range separation, robust across 8 leave-one-scale-out folds); in a synthetic machine-learning task (AI Pilot v0.2), a representation deliberately aligned with a target's mechanical structure exceeded a generic compression by a large, stable margin under matched dimension and matched predictor, replicated across two independent seeds and confirmed by a predictor-independent structural metric. The framework does not claim universal superiority of any representation family. Supported by two independent anchors (Paper A, a physical system; AI Pilot v0.2, a machine-learning system), each tested under conditions the other did not share; not a claim that any single representation property universally determines accessibility, and not a claim of cross-domain universality beyond the two systems tested.