Tests · Supreme Court · October 2026

Phosphoros called 73% of Supreme Court cases.
It never understood a word.

It only heard how each justice talked to each side: pitch, loudness, pace, interruptions, reaction time. The model was frozen before the test terms were measured, then run once.

Voice alone beats the published method.

Terms OT2024 and OT2025: 115 argued cases we had not measured when the model was frozen. Each justice’s vote is predicted on its own; a case is called for the side most justices were predicted to favor, and scored against the actual winner.

On votes, Phosphoros leads the best baseline by 6.2 points (95% interval 2.0 to 10.4) and wins in each term separately. The case-level lead (6 points) is within chance on 115 cases; the vote-level result is the claim.

The experts still win head to head.
But they skip 1 in 6 cases.

47 OT2025 cases decided after January 2026. Experts: SCOTUSblog’s argument-day recaps, read for which side the bench favored. Phosphoros: voice and the case file, no words.

Experts plus voice beat experts alone.
In both test sets.

A recap gives the bench’s mood. The voice adds each justice’s lean. The same model predicts votes from the case file and the expert read, with and without Phosphoros. This test was written down before it was run, and it passed in both sets.

These are votes, so they sit below the 84.6% case figure above: a recap gives one lean per case, and many justices break from it.

What it heard.

For each justice, each argument and each side, from that justice’s own turns. The model uses the difference between the two sides, plus how the rest of the bench behaved.

• PitchMedian pitch over voiced speech.
• LoudnessMedian level over the same speech.
• PaceWords per minute while speaking.
• Share of the floorWords and turns per minute of each side’s time.
• InterruptionsTurns that cut counsel off mid-sentence.
• Reaction timeSeconds from counsel’s last word to the justice’s first.

Data: Oyez argument audio and timed transcripts; votes from the Supreme Court Database. Trained on OT2015–OT2021, without the telephone arguments of 2020–21.

A pattern in how a justice speaks predicts the vote. It does not show why the justice voted that way.

Check it yourself.

The rules were committed to git on 4 October 2026, before any audio was downloaded. Every later test and correction is a dated section of the same file, and each test was written before it ran, including the ones that failed. This model was frozen at 21:31 UTC that night, before the first OT2024–25 file was downloaded.

• The rules and every result, uneditedEvery test, amendment, correction and failure, in the order written.
• Every number on this pageCounts, accuracies and 95% intervals, produced from the frozen results.
• Every predicted vote, OT2024–25983 votes: the model’s probability, its call, the baselines and the actual vote.

Cases are scored against the actual winner in the Supreme Court Database. Intervals come from 2,000 case-bootstrap resamples (seed 20261005).