How it works

Five AI model families ran four team configurations on real emergency cases. One setup pooled counterfactual demographic views each round without retraining.