Elements of safe and aligned Artificial General Intelligence

This dissertation studies how representations, observations, and actions determine what an observer or agent can classify, verify, and use. Its first part develops observer-relative classifications of artifacts. Structured meaning, residual opacity, and functionality are separated in the MCK framework; the resulting regions are then connected to admissible realization mechanisms and to a grammar-mediated geometry of intentional reference. A discrete encoding problem makes the representational issue concrete: a separating family stores one value per grid cell while allowing indexed tracers to recover distinct superimposed curves, and a local indistinguishability result identifies the limit of state-blind recovery. The temporal part organizes safety, cosafety, liveness, and coliveness into their intersection algebra and studies the observational mechanisms by which such properties are monitored. Touch and crossing are treated as distinct threshold modalities; bounded temporal implication is classified relative to finite observation; monitor parameters select and adapt the observation class; and a Hensel–Newton refinement construction describes qualified confirmation from sequences of finite verdicts. The final part extends these ideas to agents through typed actions with prefix-verdict semantics and an observer-relative cognitive algebra for self-attribution, identity, integrity, and freedom. The contribution of the dissertation is organizational as well as formal. It presents the deposited results in one notation-aware narrative while preserving their distinct assumptions and epistemic strength. The synthesis shows that the chapters share a recurring discipline: a claim is meaningful only relative to the representation available, the observer performing the test, and the evidence that the observation mechanism can retain. No stronger cross-family theorem is claimed.