Conscious Deployment: A Working Computational Model of Consciousness

We tested a specific account of conscious deployment by implementing its causal chain and then cutting the links one at a time. Deployment was operationalised as selective recruitment of learned capacity into an action-facing world model. In Study I, prediction error from a 15,342-parameter light transformer recruited a separately trained 4,630-parameter adapter. Across 20 confirmation seeds, recruitment raised changed-world sensory accuracy from 34.6% to 90.6% and reduced realised action regret from 1.744 to 0.531. Matched branching amplified the misspecified light model's scenario-specific guidance and worsened regret to 2.602; a post-freeze matched-action control localised the penalty to that guidance rather than to action frequency. Expert futures had no behavioural effect when their outputs were blocked from action control. Equal-count routing and cross-modal interventions further localised the effects to model-relative recruitment, relevant learned structure and governance. The frozen gate set passed 20 of 21 aggregate criteria. Study II added voluntary and insufficiency-driven opening, persistent fading episodes, consequence-learned meaning, transition learning and specialist creation within one life. Persistent control covered every required cycle at 49.3% of always-on logical cost. Learned meaning outperformed supplied and random comparators, and learned transition planning reached 98.1–100.0% accuracy. The complete frozen study nevertheless failed at 23 of 25 gates, including a revocation hand-off defect. A post-freeze repair corrected all 17 exposed hand-offs but remains exploratory. Selective operation reduced ongoing abstract cost, whereas training made the short simulated life 4.7% more expensive than always-on control. The contribution is a preregistered causal dissection of a deployment architecture, not a whole model of consciousness and not evidence of phenomenology or machine consciousness.