Chiyoda: Entropy-Guided Information Control for Hazard-Coupled Pedestrian Evacuation

Emergency evacuation is often treated as a physical routing problem in which warnings simply provide better information. The harder control problem is that the same message can improve local beliefs while synchronizing many agents toward a hazardous or capacity-constrained route.

We present Chiyoda, an Information-Theoretic Evacuation Dynamics framework that treats emergency communication as a controllable safety action in a coupled physical-information system. Agents carry probabilistic exit and hazard beliefs, exchange distorted local gossip, receive beacon and responder messages, suffer physiological impairment under CBRN-like hazards, and route through belief-weighted paths under social-force crowd dynamics.

Chiyoda evaluates static, global, responder-relay, entropy-targeted, density-aware, exposure-aware, and bottleneck-avoidance policies with telemetry that links belief entropy, belief accuracy, exposure, queueing, exit imbalance, and evacuation time. In a 50-seed primary study with 400 runs and 17780 intervention events, static_beacon has the highest information-safety efficiency in the tested scenario (0.014), while global and adaptive policies show that broad reach and entropy reduction are not sufficient safety objectives. A 900-run robustness grid further shows that static local messaging remains the most information-safety-efficient tested policy across all hazard and familiarity regimes, even though the evacuation-count winner varies. A bounded LLM extension applies the same standard to generated guidance: validated sparse messages can be highly information- efficient, but equal-budget generated guidance weakens that advantage and does not eliminate harmful convergence.

The central claim is therefore not that more information or richer language is always better, but that emergency communication must be evaluated by coupled belief and safety effects.