Data

Court opinion extractions

Structured extractions from 25 Supreme Court opinions (18 EP, 7 DP) used to illustrate and ground the model's applications.

Source PDFs

PDFs downloaded from Library of Congress U.S. Reports and supremecourt.gov. Re-downloadable via cases/download.sh. PDFs are gitignored (large binaries).

Extraction pipeline

• cases/extract_prompt.md — system prompt defining the extraction schema
• cases/extract.py — Python script calling GPT-4o API to extract structured data
• cases/extractions/*.md — one markdown file per case (25 total)

Reproducing: python cases/extract.py cases/*.pdf (requires OPENAI_API_KEY in env).

Extraction schema (per case)

1. Holding ($H_t$): quoted holding + translation to constraint on admissible $(w,c)$ + open questions
2. Fact vector $z_t$:
   • 2a. Raw salient facts (what the Court treats as legally relevant, with quotes)
   • 2b. Dimension mapping to EP/DP dimension dictionaries (D1–D8)
   • Unmapped facts flagged for potential dictionary expansion
3. Treatment of prior holdings: status (relied on / extended / distinguished / limited / overruled) + model interpretation
4. Overruling: constraint removal, justification mapped to stare decisis factors
5. Breadth: narrow reading, broad reading, breadth ambiguity
6. Concurrences / dissents: alternative constraint structures
7. Reasoning revealing implicit weights: quoted passages showing how dimensions are weighted

Notes

• Brown v. Board extraction was hand-written as the gold standard; all others generated by GPT-4o with Brown as few-shot example.
• Three long opinions (Dobbs, Parents Involved, SFFA) were truncated to fit GPT-4o's context window; concurrences/dissents may be incomplete for these.
• The project previously used the Supreme Court Database (SCDB) for an Epstein & Posner (2016) loyalty replication. That work is archived in git history.