The methods are fixed, and the pages print whatever the current file yields.
In-sample accuracy is not reported. A justice is compared with the Court they sat on. Licenses stay on the source records.
AVote and agreement statistics
In the majority, all cases, is the share of a justice's votes in the majority. The divided-case share keeps cases with at least one vote in the minority. One-vote decisions are those whose majority exceeded the minority by one.
Pairwise agreement is the share of divided cases, both justices sitting, in which the two were on the same side. Term views hide pairs under 50 shared cases. Career views hide pairs under 100.
Liberal share uses the Supreme Court Database direction code and is not shown before 1946. Each justice is compared with the Court during the terms that justice served. Cells under 25 votes are suppressed.
Single-term rates are labeled noisy.
BThe political map
The map uses published Martin-Quinn means. This build does not estimate ideology from votes. Negative is liberal. The Court median is the grey band.
Bench order uses the Martin-Quinn mean for the case's term. A score within 0.5 of zero is drawn as the center. Scores are relative to the docket.
Ideology rows in this file run through 2017 and votes through 2014. Intervals are absent until the direct Martin-Quinn pull.
CThe outcome backcast
The target is a petitioner win. For each test term T from 1950 onward, the model fits on terms T minus 25 through T minus 1 and scores term T only. In-sample accuracy is not computed and is not stored.
Layer 1 is a shrunk cell rate for the petitioner group and the respondent group: wins plus baseline times 5, divided by n plus 5. An unseen cell returns the training-window petitioner rate. The training window must hold at least 200 cases or the term is skipped.
On the published first-run file the layer 1 model calls 63.5% of outcomes and the always-petitioner baseline calls 64.0%, on 8,095 test cases. Brier scores are 0.2272 for the model and 0.2282 for the baseline. These are results, not targets.
Layers that add issue area, lower-court direction and source court are logistic regressions with L2 regularization at C = 0.5, reported from the first-run file. They are refit only when that pipeline is run on the same Supreme Court Database release.
Figures on this build are the published rolling-origin run, because the raw release is not in this repository.
Where a number on a page comes from
Court-wide counts, the backcast, calibration, win rates and the unanimity series are the published rolling-origin run in the handoff results file when the Supreme Court Database release is not on disk. Setting DATA_REPO or placing the modern and legacy justice-centered CSVs in data/raw recomputes the party-type layer and replaces those summary figures.
The October Term bench, the featured justice's career tables and the featured case narrative come from the handoff site file. The pipeline recomputes vote statistics from vote lines and uses them when they cover the service. In this fixture the vote lines cover a single case, so the career tables stay on the published aggregate and say so in the chart footnotes.
Martin-Quinn paths come from the ideology extract. They are not estimated here. Biography dates that the reference page states and the vote file does not structure live in data/fixtures/overlays.json until the Federal Judicial Center export replaces them.