Testing Two-State Access in Brain and Language Model: Human EEG Reproduction, Simulation Audit, and a Proposed Model Assay

Abstract

Brains and the hardware of artificial neural networks are both organized ordinary matter, and whether an artificial system's organization can support consciousness remains open. This paper examines one proposed mechanism, access to a capacity-limited global workspace, through one prediction of the global neuronal workspace theory: near threshold, the single-trial response distribution is a mixture of two states rather than a continuum. Two completed analyses are reported. First, the published competing-model test of that prediction is reproduced from the authors' code on the open EEG data of twenty participants. In the active session the two-state model's protected exceedance probability first exceeds 0.95 at the same 315 ms window, with a modest predictive advantage of about 0.003 nat per trial and optimizer-sensitive interval edges; in the passive session, an additional analysis, a graded comparator ranks highest without an across-window decision. Second, a held-out family comparison proposed for transferring the test to language-model representations is calibrated in simulation at one layer: it made no false two-state call in 12,000 datasets under twelve graded nulls, but its concept-cluster interval under-covered in six settings, down to 0.22, so a replacement interval must be validated before any confirmatory use. The proposed model study, with a natural-text evidence dose, a target bridge and a state-conditional causal test, is specified but not run. No claim about experience is made.