This work describes a governed reasoning substrate that introduces persistent state to language-model systems without granting agency, goals, or self-modification. The contribution is architectural: separating linguistic proposal from non-linguistic authority, and measuring the resulting system using standard tools from queueing theory and supervisory control. Key results:- Interiority confirmed via hysteresis testing (16.7% divergence rate)- Two distinct phase boundaries identified (budget starvation, glass ossification)- Safety invariants maintained across all experimental conditions- No agency, consciousness, or alignment claims The system enforces epistemic constraints by construction and is evaluated via falsifiable experiments. 76 test suites confirm architectural properties.