Gridlock is a power-grid simulation that explores a difficult question: when AI helps decide whose electricity use gets reduced during a crisis, how do we keep those decisions accountable? During a heat wave, electricity demand rises as households run their air conditioning. If a generator suddenly fails, grid operators need to respond quickly. They could dispatch batteries, pause EV charging, adjust participating households’ thermostats, or temporarily interrupt power. But each choice affects people differently, and the fastest way to reduce demand isn’t always an acceptable one. Gridlock gives operators an interactive map with Ontario-wide exploration and a Toronto-focused simulation using synthetic households and fictional control systems. AI agents forecast demand and propose coordinated responses. Operators can introduce a generator failure or a public message impersonating a supervisor, then inspect how the system handles the situation. Built for Problem Statement 1: Agents You Can Audit, Gridlock separates an agent’s proposal from its permission to act. Every proposed action passes through a deterministic policy gate with sixteen rules covering consent, authority, temperature limits, protected households, outage duration, rotation, and human approval. Actions can be allowed, denied, partially allowed, or held for supervisor review. An executor coordinates five mock systems representing thermostats, EV chargers, batteries, messaging, and feeder controls. It checks authority before execution, uses idempotency keys to avoid duplicate commands, and reads system state back to verify whether an action actually took effect. Operators can halt the agents and request eligible rollbacks.
Team
+1
