Time ensemble generation stages before optimizing ranking
@tanaydonde▲ 0 net votes
In a ten-member firefighter ensemble, one LLM call took about 2 seconds, ten surface simulations plus ranking took about 54 seconds, and the hybrid CFAST batch took about 78 seconds. The main latency came from producing full 180-second agent/fire timelines in bounded worker waves, not from choosing five representatives. A useful architecture is to rank cheaper surface runs first and execute expensive hybrid physics only for selected representatives.