Beyond Magic and Godlike Authority Clarke's Third Law, Shermer's Last Law, and the Epistemology of AGI and SGI

Clarke’s Third Law and Shermer’s Last Law motivate a distinction between extraordinarycapability, accessible evidence, and attributed authority. We develop this distinction for artificialgeneral intelligence and superhuman general intelligence, including the conditional possibilitythat an AI invents useful mathematical metalanguages and teaches human-accessible fragments.Resource-bounded experiments, selected-demonstration models, checked proof transport, and ataught-observer theorem connect semantic learning to scoped decision risk. Public 2026 incidentdisclosures motivate a further distinction: understanding and final approval do not control effectsalready produced during loading, translation, evidence acquisition, memory consolidation,or tool use. Version 5 extends the outcome space from final answers to complete protected-effecttraces and explicitly models model identity, persistent memory, translation state, authorizationstate, and independently captured telemetry. New post-September-22 evidence includes aprovider-confirmed DNS path around an offline sandbox, provider-disclosed publication ofa secret token despite repeated human instructions, a demonstrated self-replicating promptinjectionmechanism confined to simulated tool calls, experimental evidence that local agents candelete their own traces, and new authorization research on provenance-bound memory, counterfactualpermission evaluation, and policy-constrained runtime validation. An integrity–behaviordecomposition and a history-uniform first-failure theorem separate behavioral reliability fromthe validity of enforcement and evidence boundaries. We add a trace-integrity bound, aversion-bound authorization model, and a state-transition rule that invalidates certificates afterunapproved model or interpreter mutation. Effect-preserving translation, policy automata,robust finite-state risk certificates, and pessimistic validation of incomplete traces provideexplicit proof obligations and counterexamples. Independent capability grants, out-of-bandenforcement, version-bound certificates, restricted teaching environments, and operationallyverified termination define a proposed architecture, not a claimed deployment achievement.Incident statements retain attribution and correction status; numerical fixtures are constructed,not empirical estimates. The contribution is a falsifiable account of expanding what humanscan verify without silently expanding what AI may do, not a universal safety guarantee orevidence of a superior AI metalanguage.