Measuring Functional Access in a Language Model by Intervention: PCI_T and FCA-Bench

Whether the information held by a language model is genuinely accessible—whether it propagates in an organized way and whether the system can govern it—cannot be read from its answers or from its evaluation scores. We present two instruments, built independently of one another, that interrogate the system by intervention rather than by description. PCI_T transposes to transformers the logic of the perturbational complexity index used in clinical neuroscience: perturb an internal state and measure the complexity of the response that propagates from it. An adversarial bench shows that absolute complexity is not sufficient—random or chaotic systems produce as much as organized ones—but that its directional specificity separates them sharply: an enrichment close to 1 for adversaries, and between 7.1 and 8.4 for structured systems. On Qwen3-8B this specificity is localized late in the network; frozen and then replicated on an independent holdout, it yields a final measure of 1.381 (95% CI [1.201, 1.609]): the endogenous direction of the residual state at layer 27 produces about 38% more complexity than orthogonal directions of the same norm. FCA-Bench evaluates functional access along two profiles that are never aggregated: a behavioral profile, observable through the ordinary interface, and a causal profile, which requires an architectural primitive or an exposed intervention. On standard Qwen3-8B (twenty instances per test), the behavioral profile is contrasted—timeline reconstruction 0.89, temporal self-localization 0.85, duration tracking 0.26, controlled erasure 0.72—and five of the six causal tests receive the status of structural absence, in line with the preregistered expectation for a standard transformer. The two instruments thus provide a measured baseline for the controlled comparison they are designed for: the same model equipped with a latent loop.