A2ZAI LAB — POPULATION TRIALS
First the pool earns the right to predict
A pool of agents must reproduce a real flu season it never saw — scored against the actual reported curve — and publish its calibration score. Below 70, the failure publishes instead of conclusions, and every what-if stays locked.
Calibration scores so far — Texas, fall 2025 (threshold: 70)
v0 — rule-based pool, 1,000 agents19.5 / 100
v1 — Muse Spark planner, 100 agents21.1 / 100
v2 — Muse Spark planner, 100 agents17.2 / 100
v2 — Muse Spark planner, 1,000 agents17.3 / 100
Every run so far has failed the gate — a peak-timing miss of 3+ weeks caps the score at 49. The console below replays the latest run in full: the agents, the curve it produced, and exactly where it diverged from reality.