A Neurosemantic Communication Layer Between Human and AI for Autonomous Materials Science: Biological State, Dynamic Coupling, and Epistemic Integrity

Autonomous materials systems increasingly operate as closed-loop agents in which materialstate, experimental state, model uncertainty, and candidate actions are represented dynamically,while the human participant still enters mainly through commands, annotations, approvals, orother discrete interventions. We develop a joint architecture in which continuous human biologicaldynamics can become an internal state variable of the human-AI system while remaining subject toindependent epistemic and authority constraints. The human-state layer is staged: longitudinalEEG -> criterion-bound state reduction -> recurrent structures -> independent functionalgrounding -> operational neurosemantic units -> a continuously updated Ht representation.Recurrence alone does not confer semantic status. The machine-side integrity layer separatelyrepresents local integrity, independent corroboration, external reality consistency, temporaltrust, persistence, and authority. Candidate states progress through TRANSIENT -> WORKING-> CONSOLIDATED, but consolidation does not self-authorize memory modification or physicalaction.We report three computational analyses. First, a synthetic integrity benchmark comparesthe separated-gate architecture with coherence-only, source-count, scalar-trust, and change-pointbaselines. In addition to legitimate recurrence, noise, contradiction, coordinated false recurrence,and genuine change, we add two adversarial extensions: a false signal that imitates the naturalnessstatistics used by the corroboration channel, and a legitimate common-mode signal that depressesthat channel despite external consistency. The full architecture rejects both coordinated-falseregimes because persistence still requires external Reality confirmation, but it also fails toconsolidate the legitimate common-mode condition under the present hard corroboration gate,exposing a concrete false-negative failure mode. Operational-threshold sensitivity remains perfectthrough ±20% perturbation for the primary R1, R4a, and R5 regimes; R1 begins to deteriorateonly at ±30% in the specified sensitivity procedure. Second, we audit the previously reported human-state coupling experiment. Once no-Hpolicies receive the same defer action and matched decision thresholds, the apparent advantage ofthe original H-aware policy largely disappears. Mean normalized cost is 0.8403 for a no-H deferpolicy at the matched 0.52 threshold and 0.8428 for the H-aware matched policy; the paireddifference is 0.00245 with an empirical 95% interval of [−0.00119, 0.00718]. A no-H defer policyat 0.40 achieves 0.6256, showing that the original gain was primarily an action-set/thresholdeffect rather than evidence for the value of Ht. Third, a decision-theoretic value-of-information analysis gives all policies the same actionset and evaluates the incremental value of observing the binary human state. In the originalsynthetic world, perfect Ht reduces optimal expected cost by only 0.00099, and this valuedecreases monotonically toward zero as observation noise increases. Across a broad parametergrid, however, the value of Ht ranges from zero to 0.182, identifying the conditions under whichhuman-state information can genuinely change the optimal supervisory action. These resultsvalidate architectural distinctions and expose their limits; they do not constitute biologicaldecoding. Prospective longitudinal EEG, held-out temporal replication, functional grounding,and comparison against strong non-EEG baselines remain required before the neurosemanticlayer can be treated as empirically established.