From Viable Continuation to AI Governance: An Explicit Axiom Ledger and Limits of Model-Relative Guarantees

Can scientific models of viable continuation constrain artificial-intelligence governance without quietly converting evolutionary persistence into moral authority? We propose an evidence-qualified audit architecture separating four layers: modeled consequences for affected systems, explicit normative premises, epistemic qualification of the causal model, and accountable authorization. Within a finite, shared continuation language, viability-option inclusion defines an observer-relative consequence preorder. We formalize four boundaries: weak dominance is preserved when affected-observer sets contract, while strictness may first arise when an observer is added; symmetric conflicts cannot admit a unique deterministic choice under both stakeholder anonymity and action neutrality; a harm certificate is sound only if the actual environment is represented by its certified model set; and observationally indistinguishable environments with conflicting outcomes cannot be safely separated by an observation-only detector. These propositions are elementary, model-relative and not claimed as new theorems of social choice or verification. Their use is methodological: they motivate a typed decision record retaining omitted stakeholders, uncertainty, unresolved conflicts, external provenance of objectives and the source of normative authorization. We provide a finite countermodel verification suite and falsifiable experimental studies. The framework does not derive ethical obligations from physics, demonstrate emergent morality, validate AI regulation, or identify possible machine consciousness.