Thousands of computer models simmer in a single cauldron, and the steam rising from them is thick and impenetrable. Each whispers something about the future of El Niño, the warm current that, every few years, reshapes climate across the planet. But in the general clamor, you cannot tell who is right and who is just blowing hot air. The scientists devised a still: they keep only those models that have already nailed the event's phase and distill them down to twelve pure drops. These drops are the typical states that successful predictors converge on.

The distillation is straightforward. First, fifty models are run on satellite data: sea surface temperature, temperature differences with depth, and wind stress. Each forecast month gets its own pot. Then comes the tasting: a model that got the phase wrong is poured away, while one that got it right is kept. The surviving models are compressed into twelve states, but not arbitrarily; the compression respects which features each model deems important. The result is a map: you can see what the model is looking at when it makes its prediction.

Previously, such maps were drawn after the forecast was made, and they often wobbled. Now transparency is built into the distillation process itself. And if a drop turns cloudy, you can see which feature got stirred up.