Regulatory Ground for Agentic AI — Band-Limited Optimization for Trustworthy Discovery

This working paper proposes a practical thesis for agentic AI and robotics: trustworthy discovery requires regulation. As systems move from “oracle” task competence to tool-using and embodied agency, failures shift from isolated errors to cascades (tool misuse, unsafe exploration, goal drift, runaway optimization). The paper defines a concrete Regulator Spec (risk bands, hard constraints, budgets, uncertainty gating, checkpoint/rollback) and illustrates two instantiations: (A) a tool-using research agent and (B) a mobile manipulator using runtime assurance (RTA) switching with a conservative safe mode. It also distinguishes task competence from epistemic agency (exogenous vs. endogenous revision) and outlines minimal tests for contradiction-driven model revision. Newer revisions clarify telemetry→posture mapping via a global restraint signal and treat learned judges/critics as telemetry (not authorities), using abstain/tie mass as a first-class uncertainty signal that tightens posture in verifier-free regimes. Status: draft / request for critique and pointers to prior art.