Methodology

A true jury of the public: real people, on the record, tested against what everyone assumes juries do. Every case we run is an experiment in how the public actually judges.

One verdict per visitor per case, deduplicated by an anonymous browser token. Counts display as percentages until a case reaches 150 verdicts, then raw counts publish alongside. Human tallies contain human votes only; the AI Jury Forecast on each case is clearly labeled and never counted.

Some cases secretly assign each visitor one of two case files. Nothing in either file is false; they differ only in how much of the record they contain. The split stays sealed until the case is decided, and the verdict gap between the two juries publishes as a finding with sample sizes attached.

Cases marked Forecast concern matters pending before real courts. Our numbers there are the public jury's assessment of the public record and a prediction of the outcome, never a judgment of the real case. Forecast jurors hold a standing verdict they may change as the trial unfolds; every change is recorded, never overwritten. The published flip rate counts jurors who changed their verdict at least once. Because we re-ask the same jurors rather than sampling new people, sentiment movement reflects changed minds, which distinguishes the tracker from polls and prediction markets. On forecast cases, the official forecast locks as of the moment the case is submitted to the real jury, reconstructed from the timestamped record; voting may continue through deliberations, and those later votes are reported separately.

Our tally is a first ballot. Real juries take one too: an initial vote before deliberation begins, followed by hours or days of argument aimed at unanimity. We do not simulate that argument, and we will not pretend to. What we publish is the distribution of independent judgments on the record, before anyone has been pressured toward agreement.

An objection worth raising against us: a real jury never returns a percentage. It convicts, acquits, or hangs. So why would a first ballot forecast anything? Because the central finding of the field says it does. Kalven and Zeisel's The American Jury, the study modern jury research is built on, found that the side ahead on the first ballot wins the deliberated verdict roughly nine times in ten. Deliberation rarely reverses a room; it converts the majority into unanimity. Later work has kept finding the same pattern. A first ballot is not a weaker version of a verdict. It is the input that best predicts one, captured before conformity pressure gets to work on it.

The honest consequences, which we accept: a near-even first ballot forecasts a hung jury or a knife-edge outcome, not a winner, and we report it that way. And our jurors vote independently rather than in one room, so we forecast the direction a deliberating jury starts from, never the theater of the room itself. Our published resolution rules score direction, which is the thing we actually measure.

A prediction market aggregates the opinions of people betting on an outcome, most of whom have read the coverage, the odds, and each other. It is a powerful instrument for measuring expectation. It cannot tell you why anyone believes what they believe, its participants rarely sit with the primary record, and its price moves when the crowd's mood moves, whether or not any evidence changed.

Every number here comes from a person who was first handed the record. Each verdict can carry written reasoning, entered into a public archive. When our line moves, it is the same jurors changing their minds, on the record, not a changed sample or a shifted mood, and the flip rate says exactly how many moved. Our forecast locks the moment the case goes to the real jury, its sealed value is committed by published hash before the outcome is known, and the rule for scoring every possible return is published in advance, including the returns we score as a void. A market tells you what the crowd expects. This instrument tells you what informed people concluded from the evidence, one by one, why, and whether they held to it, under scoring rules fixed before the answer existed. Those are different measurements, and for understanding how juries and publics actually judge, the second is the one that matters.

Anyone, anywhere, may serve. That is by design. The questions our cases ask are human questions before they are national ones, and a jury drawn from everywhere tests them the way no courtroom can. We therefore do not claim our verdicts represent any one country. When jurors choose to share where they serve from, comparative verdicts across places will publish as labeled research findings; until then, the tally is exactly what it says: the public, counted one juror at a time.

Every case file is drawn from the public record and lists its sources on the case page. Court records and named outlets, checkable in one click.

This methodology is versioned; changes are logged on this page. Our accuracy record, including our misses, publishes in the press room.