Computable Design Space Exploration via Governance Frameworks: Joint Optimisation of Permission Boundaries, Monitoring Thresholds and User Transparency Policies (for Proxy AI)

Agent-based AI embeds language models within tool invocations, memory read/write operations, and multi-agent delegation chains. Consequently, risks shift from ‘erroneous output content ’ to ‘expanded action spaces, delegation topology, and interaction drift’—triggering unauthorised access, irreversible operations, and accountability breakdowns. This paper proposes Computable Governance Design Space Search (CG-DSS): formalising governance artefacts—authority boundaries, monitoring/escalation thresholds, and user transparency policies—into a computable hybrid design space. Under hard constraints, it performs constrained multi-objective Bayesian optimisation with verification filtering, yielding provably feasible Pareto governance sets. To support reproducible and auditable evaluation, this paper constructs a NIST-grade trajectory- level hyper-simulation data protocol aligned with the NIST AI RMF and GenAI Profile's ‘governance-measurement-management ’ closed loop. Each sample constitutes a replayable audit trail, employing audit fields consistent with the intent of NIST SP 800-53 Rev.5 AU-3 to ensure traceable evidence chains. Under the provided six-level fine-grained cognitive map (Nodes/Edges/Triples) structural prior, the authors generated 600 auditable trajectories and 7, 150 event-level audit records, comparing multiple governance strategies under a budget constraint B=60. Results demonstrate: CG-DSS maintains 100% feasibility (validation pass) while significantly reducing residual risk and sustaining stability during drift segments. In contrast, ‘pure optimisation ’ neglecting validation filtering achieves greater Pareto supervolume but at substantially higher infeasibility rates, with failures predominantly occurring in localisable mechanism chains such as monitoring boundary breaches. This paper provides a complete dataset and auditable schema, rendering conclusions falsifiable, reproducible, and auditable.