Multi-Agent Simulation as Qualitative Predictive Gating Layer.

Sandbox simulates counterparty reactions before any sensitive action dispatches. 25 claims.

Before dispatching any typed-action invocation classified at or above a configured sensitivity tier, the runtime spawns a bounded multi-agent simulation in a sandbox. A scenario builder constructs a sandboxed environment populated with simulated counterparty agents representing the recipients, downstream services, regulators, and the user-future-self that the action would touch. The proposed action is executed inside the sandbox; the simulated counterparties produce reactions; a qualitative outcome classifier scores the reactions across declared dimensions of privacy harm, financial harm, reputational harm, regulatory exposure, and reversibility; a gating engine refuses, downgrades, or permits the action against a configurable threshold. The simulation transcript and outcome score are signed and recorded as a typed SIMULATION-DECISION event in the decision-lineage ledger. Filed 4 May 2026 as GB2610422.4.