We propose a formal paradigm of ontological engineering for Artificial General Intelligence (AGI) based on the Conflict–Moment–Impulse (CMI) theory. Ethical invariants are encoded as a preserved G-functional, ensuring alignment not through external constraints but through the system’s internal dynamics. We formalize partial ethical autonegation for gray-zone dilemmas and prove a theorem of non-removable constraints. Implementation pathways include neuromorphic embeddings, cryptographic commitments, and evolutionary design. This approach reframes AGI safety as an ontological firewall: crises are processed through CMIresets that preserve invariants while enabling adaptation.