EMBERBench methodology

EMBERBench evaluates step-by-step decisions in simulated fire evacuations. Each model observes the current world and chooses how people move. Recorded runs are replayed and verified before safety and execution quality are scored.

Transparent scoring

EMBER is a family-balanced composite out of 100, not a survival percentage. The public scoring contract specifies its formula, weights, gates, aggregation and failure handling. All models in a comparison use the same recorded environments.

This website reports public-development evidence. Models may train or tune against it. The current snapshot has one trial per model; it does not establish a statistically significant winner or certify real-world evacuation safety.

Read the complete scoring contract, replay verification, statistical methods and evaluation boundary. The interactive reading view becomes available when its data loads.