Anantam

EMBERBench

Anantam Labs’ first contribution: a benchmark for how AI guides people through simulated fire evacuations, where routes can close and queues can form.

At each step, a model sees the current situation and chooses how people move as the fire spreads. It must respond to blocked paths and congestion without knowing where the fire will spread next.

Models are evaluated in the same frozen environments. Each recorded run is replayed and checked before scoring for safety and execution quality. The published results cover the public development suite, which models may train or tune against. They are reproducible development evidence, not contamination-resistant hidden-test results.