ResearchPacks
Cryptographically sealed, self-contained evidence bundles that eliminate external database dependencies at runtime.
The Reproducibility Problem
Traditional benchmarks evaluate models against live external datasets. When those datasets change, scores become incomparable. When databases go offline, evaluations break. To reproduce coverage numbers from six months ago often requires complex database reconstruction.
LegalChain solves this by sealing inputs at build time. A ResearchPack travels with its evaluation instance, guaranteeing that the inputs used to produce a score today are identical to those used next year.
Anatomy of a ResearchPack
Anchor Opinion
The full, unaltered text of the Supreme Court majority opinion under analysis.
Cited Authorities
Authoritative summaries (Syllabi/Head Matter) of the top-ranked precedents cited by the anchor.
Citation Evidence
Context snippets, treatment signals ("followed", "distinguished"), and metadata.
Deterministic Ranking
To select the most relevant authorities, we use a deterministic formula combining multiple signals:
Scoring Factors
- Log-scaled Citation Frequency
- Shepard's Treatment (Followed/Criticized)
- Introductory Signals ("see", "cf.")
- Fowler Precedential Score
Optimized Summarization
Including full text for every cited case would explode the context window (>400k characters). Instead, we use Authoritative Summaries:
SCOTUS Syllabi
Official summaries prepared by the Reporter of Decisions. High-quality, court-approved abstracts.
CAP Head Matter
Editorial summaries and headnotes from published reporters, widely used by legal professionals.