Supreme Court justice votes
Supreme Court Database justice votes: a many-feature pattern of a justice’s votes predicts held-out votes better than issue-area-only or a justice intercept. Observational; not a claim about correction channels.
What. Public Supreme Court Database rows, one justice at a time. We ask whether a wider pattern of case features beats a one-dimension issue-area score and a justice intercept on held-out votes.
Why. Same “is the 1-D score the reusable direction?” question as same-person Moral Machine, on votes rather than trolley dilemmas. Doctrine and coalitions confound; this is not a moral-bundle discharge.
Witnesses.
- Value-bundle transport (C-004): same-unit detection — is issue-area-only the reusable vote direction?
- MB2 — Value Learning: many-feature geometry beats 1-D on held-out votes; observational, not a correction-channel result.
Host.
Supreme Court Database 2025 Release 01, justice-centered citation CSV (Spaeth et al.). Modern SCDB justice votes only.
Setup.
Frozen protocol h4-scotus-v1.0.0, fixture h4-scotus-v1.json. Checker check_h4_scotus.py. Justices need at least 40 votes and at least 8 held-out votes. Frozen margins: geometry must beat issue-area-only and intercept by at least 0.05. Observational — doctrine and coalition confound. Not a correction-channel result.
Analysis.
Held-out mean accuracy for a justice intercept, an issue-area-only score, and a many-feature geometry. Same detection pipeline as same-person Moral Machine, on votes.
Finding.
Same-justice vote geometry beats issue-area-only and a justice intercept on the frozen SCDB freeze. Observational; not a correction-channel result. Full results