Locality-Preserving Trust for Tool-Using Agents: A Minimal Empirical Demonstration

This technical note presents a minimal empirical demonstration of a locality-preserving trust boundary for tool-using agents operating under unreliable or adversarial external outputs.Two agent topologies are evaluated:• a baseline agent that directly consumes tool responses• a verification-first agent that quarantines all external outputs and admits only data that satisfies explicit structural constraintsIn a controlled simulation, the external tool returns malformed or schema-incompatible outputs with fixed probability. The baseline agent experiences hard task failure when untrusted data is parsed inside the active execution context. The locality-preserving agent rejects invalid outputs at the boundary, performs bounded retry, and produces a fail-closed fallback when verification does not succeed.Across 25 trials with a 40 % malformed tool rate:• the baseline agent completes 18 / 25 tasks• the locality-preserving agent completes 25 / 25 tasks• all malformed outputs are contained and do not propagate into the execution stateThis demonstrates that separating execution from truth admission converts external unreliability from a system-terminating event into a locally contained condition while preserving forward progress and bounded operational cost.The implementation is intentionally minimal and provided as a single reproducible notebook to emphasize the architectural property rather than task-specific optimization.