Legal Digital Twins
Predictions on the record before the Court hears the case
October Term 2026
Raymond Duch · Laurenz Günther
Nine digital twins — one for each sitting Justice of the United States Supreme Court — each conditioned only on that Justice’s own authored and joined opinions from October Term 1991 to October Term 2024. For each new case a twin is given a one-page description prepared from the Court’s question presented and the parties’ briefs, together with eight passages retrieved from the Justice’s prior writing, and asked how that Justice will vote and why. Nothing from the current Term is in its evidence; the case under prediction is excluded from retrieval; every run is logged with the hash of the case description, the retrieved passages, the prompt template and the model identifier returned by the API.
Predictions are filed on FantasySCOTUS before argument, published in a public repository the same day, and never edited afterwards. The decisions arrive months later. That gap is the point: it is the only test of a prediction method that cannot be argued with.
October 2026 sitting — batch 1, filed 4 October 2026
| Docket | Case | Argued | Twins’ prediction |
|---|---|---|---|
| 25-170 | Suncor Energy v. County Commissioners of Boulder County | 5 Oct 2026 | Affirm, 5–3 (Justice Alito recused) |
| 25-735 | Johnson v. United States Congress | 5 Oct 2026 | Affirm, 5–4 |
| 25-498 | Anderson v. Intel Corp. Investment Policy Committee | 6 Oct 2026 | Reverse, 7–2 |
| 25-579 | Department of the Air Force v. Prutehi Guahan | 7 Oct 2026 | Reverse, 7–2 |
Per-Justice votes, the probabilities each twin attached to them, the four case descriptions, the scoring rule and the hashes of the sealed inputs are in the public record: github.com/rayduch/legal-digital-twins-record (release october-2026-batch-1, 4 October 2026).
October 2026 sitting — batch 2, run 11 October 2026, not filed
| Docket | Case | Argued | Twins’ prediction |
|---|---|---|---|
| 25-5343 | Beaird v. United States | 13 Oct 2026 | Vacate (petitioner), 9–0 — not filed, unscored |
| 25-459 | Salazar v. Paramount Global | 14 Oct 2026 | Reverse, 7–2 — not filed, unscored |
These two cases were run on 11 October 2026, before argument, but the predictions were not entered on FantasySCOTUS before the arguments on 13 and 14 October: the filing step was missed while the project lead was travelling. Under rule 1 of the published scoring rule an unfiled sitting is not scored. The runs are published in the record for transparency, labelled as run on 11 October and published on 21 October, and are excluded from the scored record; the note in the repository says what happened and what changes. Release october-2026-batch-2-unfiled.
November 2026 sitting — batch 3, filed 31 October 2026, before argument
| Docket | Case | Argued | Twins’ prediction |
|---|---|---|---|
| 25-5930 | Maxwell v. Dinis | 2 Nov 2026 | Reverse (vacate), 7–2 — filed 31 Oct 2026 |
| 25-352 | Younge v. Fulton Judicial Circuit District Attorney’s Office | 2 Nov 2026 | Reverse, 9–0 — filed 31 Oct 2026 |
| 25-581 | St. Mary Catholic Parish v. Roy | 3 Nov 2026 | Reverse, 6–3 — filed 31 Oct 2026 |
| 25-1003 | Guerrero v. Johnson | 4 Nov 2026 | Reverse, 6–3 — filed 31 Oct 2026 |
Four cases were run on 30 October 2026 and entered on FantasySCOTUS on 31 October, two days before the first argument, each case saved once and not edited after. In Maxwell the United States, though the nominal respondent, supports vacatur and the Court appointed an amicus to defend the judgment, so petitioner means vacate; in Guerrero the State is the petitioner. The crowd at filing called all four the same way (8–1, 6–3, 6–3, 5–4). Record, packets, sealed hashes and filing screenshots: release november-2026-batch-3 in the repository above.
Next filings: Nielsen v. Watanabe and Hoffmann v. WBI Energy Transmission (argued 9 November) and Department of Labor v. Sun Valley Orchards (10 November); packets by 5 November, run 7 November, filed 8 November.
What we know so far
On the 58 argued cases of October Term 2025, with the decisions held out as the answer key, the twins matched 87.5% of individual Justice votes (448 of 512) and 52 of 57 outcomes. Those decisions may fall inside the model’s training window, so we report that figure as recovery of known outcomes, not as forecasting skill. The cases above are the test.
Read more
- Essay: Nine readers — talkingtomachines.substack.com/p/nine-readers (5 October 2026)
- Record: github.com/rayduch/legal-digital-twins-record
- Workshop: The Silicon Audience, LLMSS 2026 pre-conference workshop, City University of Hong Kong, 13 October 2026
Team: Raymond Duch (Nuffield College, Oxford) and Laurenz Günther (Toulouse School of Economics), with the Talking to Machines team. Research commentary, not legal advice.