Source: cases/SELECTION.md
# How the 60 cases were chosen `docroot/statute_cited_by.json` holds only the 50 most recent citing cases per statute, so the candidate pool was built from the IR `*.refs.jsonl` out_links of the federal reporters (us, f2d, f3d, f4th, f-supp, f-supp-2d, f-supp-3d): cases whose out_links cite the statute or a subsection of it. Per statute, all U.S. Reports candidates plus a seeded random sample of 450 others were ranked by density of element keywords in the opinion text (cf. `tools/case_pick.py`). Picked by hand from the top of each ranking: ~4 Supreme Court opinions that construe an element, plus ~8 lower-court opinions spread over four eras (<1990, 1990s, 2000s, 2010+), mixing circuit and district courts, civil and criminal, and claimant wins and losses; text length 9k–58k chars preferred (IR text is capped at 60k). The case list is the set of `.case` files in each directory. Known bias: keyword-dense opinions over-represent pleading-stage dismissals.