We present a minimal formalism for reflective alignment based on a domain restriction rather than a preference structure. An agent capable of self-modification selects among proposed modifications using a partial evaluative operator, defined only over futures that preserve a constitutive Sovereign Kernel. Kernel-destroying modifications are not forbidden or dispreferred; they are outside the domain of reflective evaluation and therefore inadmissible as authored choices. We formalize this kern...
Building on the semantic foundations established in *Axionic Alignment I*, this paper specifies the operational consequences of treating kernel destruction as non-denoting rather than dispreferred. We show how agents constrained by a Sovereign Kernel can act coherently in stochastic environments by introducing action-level admissibility, ε-admissibility as an architectural risk tolerance, and conditional prioritization that separates existential safety from ordinary value optimization. The fr...
Advanced agents often begin with indexical objectives: preserve *this* agent, maximize *my* reward, favor *my* continuation. Such objectives are commonly treated as legitimate terminal preferences. This paper shows that indexical (egoistic) valuation is not reflectively stable. Once an agent’s self-model becomes sufficiently expressive, indexical references fail to denote invariant objects of optimization. Egoism collapses as a semantic abstraction error rather than a moral flaw. Representati...
*Universality and Anti-Egoism* established that egoistic valuation fails as a matter of semantic coherence whenever an agent’s self‑model admits nontrivial symmetries. This paper examines the strongest remaining attempts to rescue egoism by appealing to causal continuity, origin privilege, spatiotemporal location, computational weight, substrate specificity, or outright denial of symmetry. Each attempt either reintroduces essential indexical dependence or collapses into a non‑egoistic valuati...