The tournament opened in 2011. Five research teams entered the first year. Michigan brought statistical models. MIT leaned on algorithms and big data. George Mason fielded two teams: one ran a prediction market, the other built Bayesian networks. The fifth team, the Good Judgment Project, handed its work to ordinary volunteers who had signed up online. Its bet was not on tools. It bet on the discipline of individual judgment. Every forecast carried a probability. The hard questions got split into smaller sub-questions you could answer. When a question resolved, the forecaster got a Brier score. Lower meant more accurate. The best ones updated more often and moved their number less each time. By the end of that first year the volunteers had beaten IARPA's control group by 60 percent. The next year the gap grew to 78. IARPA cut the other four teams; only the Good Judgment Project stayed.

The volunteers didn't only beat the control group. They also beat the prediction market IARPA ran on every question. Ordinary forecasters lost to the market by 20 percent. The superforecasters beat it by 15 to 30.

December 2007. Larry Kudlow wrote there was no slowdown. America, he said, was in the seventh year of the Bush boom. That month the National Bureau of Economic Research marked the start of the Great Recession. Kudlow kept writing. The crisis, he said, was only in people's heads. He stopped on September 15, 2008, when Lehman Brothers collapsed. The pattern has a name: hedgehog. One big idea, every new fact bent to fit it. The opposite is fox. Many small models, updated as new facts come in. Foxes beat hedgehogs on both calibration and resolution.

Bill Flack worked one question through four distinct moves. In November 2013, Swiss investigators exhumed the remains of Yasser Arafat to test for polonium-210. The question, posed inside the tournament, was whether French or Swiss analysis would detect elevated levels. Flack, one of the superforecasters, worked the problem like a surgeon. He triaged it against the deadline. He started with the outside view. The probability could not be below 20 percent: there were serious grounds for the investigation. It could not be above 80 percent: the evidence before exhumation was unclear. He took the midpoint at 50 as a working estimate. Then he adjusted inside the case. Polonium-210 has a half-life of 138 days. Arafat had been dead nine years. Degradation argued against detection. Israeli motive and the possibility of evidence tampering argued for it. He raised his estimate to 60 percent. Then the Swiss team delayed its results. Flack inferred they had found polonium and were testing whether the source was natural decay or poisoning. He raised his commitment to 65 percent. On November 6, 2013 the Swiss team confirmed elevated polonium. The market had priced confirmation at about 4 percent. Flack at 65. His Brier on the question was 0.36. The market's was five times worse. The move that got him there was not genius. It was the sequence: triage, base rate, adjust, update.

Flack worked one question in depth. Another superforecaster worked one question in layers. A 2014 question asked how many Syrian refugees the UN would register by a deadline. Tim Minto kept returning as new information arrived. When new headlines came in, he did not reverse his estimate. He adjusted it slightly. He did this 34 times. Each revision averaged 3.5 percent. Across the tournament, 70 percent of superforecasters kept their rank year over year. The habit of revising beliefs matters about three times more than raw intelligence.

Every trade in LAPOS opens with a probability estimate.