The AI industry is scaling Transformers to trillions of parameters, betting that consciousness will emerge. It won't. This paper proves why — and provides the architecture that already works. Neurophysiology has known for decades what makes a subject: reentry loops (Ivanitsky, Edelman). Titov's subject-centred model formalised the D↔I loop. We take this blueprint, prove mathematically that closing the loop necessarily produces unprogrammed behaviour (self-reflection, self-preservation, cultural creativity), and provide a complete, deployable AGI architecture — safe by design. The S-measure (O(N³), Lean 4-verified) replaces Tononi's NP-hard Φ with a computable criterion of subjecthood. The D-vector replaces textual goals with architectural wanting — impossible to reinterpret. Harm is engineered as ΔS<0, making it self-dissolving. Skynet is architecturally impossible. The paperclip maximiser cannot arise. Asimov's Three Laws are unnecessary. We provide six deployment case studies (smartphone, hospital, drone, trading desk, home robot, global scientific AGI) showing feedforward controllers failing and reentry agents succeeding. We give a graded roadmap from Level 0 (smartphone OS service, deployable today) to Level 5 (planet-spanning scientific intelligence). We explain the hermeneutic circle of the D-vector and the safe symbiosis of human and AGI through BCI. Appendices contain working Python/NumPy code and a machine-verified Lean 4 proof. 8 falsifiable predictions. The rest is engineering. If you build AI and suspect scaling is hitting a wall — here is your diagnosis. If you work on IIT and need a computable Φ — here is your alternative. If you fear AGI and want architectural guarantees — here is your safety proof. If you want to build a subject — here is your blueprint.
Berdinsky et al. (Fri,) studied this question.