Empirical studies of published judicial opinions, mostly Florida, mostly estate and trust law.
Run by Kevin D. Klagge, Esq. · stepuplaw.com ·
ORCID 0009-0002-1385-8498 ·
Start with HANDOFF.md if you are picking this up cold, then STATE.md for where the project stands, what has been run, and what is next.
Read METHODOLOGY.md before starting a study, and CORPUS-LINGUISTICS.md if the question is about what words mean or which words recur. It names the field, the standard method paper, the reporting conventions, and the two threats that most often invalidate this kind of work.
| Study | Unit | Size | What it measures |
|---|---|---|---|
disputed-terms |
(term, decision) | 286,846 | which drafted words get litigated for ambiguity, across all 51 state appellate jurisdictions, and how often the challenge succeeds |
reformation-fl |
decision | 932 | Florida reformation of wills, trusts, deeds, contracts and policies, 1853–2026, across three statutory regimes with dated breaks |
statutory-staleness |
(decision, section) | 19,085 | how much Florida case law construes a statute that has since changed — the gap a citator does not cover |
ambiguity-pools |
decision | 6,523 | the feasibility study that produced the method, including the extraction that failed and why |
notebooks/ holds the executed analyses, outputs included, so the
result is visible without running anything. They also run anywhere: where the
local build is absent, the loader falls back to the published dataset, so the
Colab button is not a link to a stack trace.
protocols/ one protocol per study, pre-registered before analysis
studies/ one directory per study: build script, report, run manifest
notebooks/ executed Jupyter analyses, outputs kept
lib/ shared classifiers and helpers
scripts/ the publishing pipeline
dist/ exported dataset metadata; the data files themselves are not in git
Queries go through ~/caselaw/, the shared corpus pipeline, never by
opening us-opinions.db directly. That library owns the court-scope
definitions, the ?immutable=1 open and the block decoder, each of which has
been gotten wrong before by someone reinventing it. Its fl scope also resolves
to a local SSD slice, so most queries return in well under a second.
Set CASELAW_HOME if the pipeline is not at ~/caselaw, and US_LAW_DB if the
statutes database is elsewhere.
⚠ The ?immutable=1 rule is for the opinion corpus, which is never written
during a query. Do not use it on us-law.db while a statute crawl is
running: the reader will report database disk image is malformed on a database
that is perfectly healthy.
Every dataset here is derived, and every one is rebuilt by a script in this repository. Git holds the code and the metadata that describes the data — Frictionless datapackage, schema.org JSON-LD, MLCommons Croissant, with a SHA-256 and a row count on every file. The data itself is published separately as a citable dataset, which is also the form to cite.
PUBLISHING.md sets out where each piece lives and how the
records reference one another, and DISTRIBUTION.md is the
running checklist of which platforms are done and which are pending. Note that the research landing pages and the DOIs
are not live yet; the metadata files name their intended locations, and
identifier is left empty rather than filled with a placeholder DOI.
These studies are not about estate tax, and the first one was living in the estate tax dataset repo by accident. A study is a different kind of artifact from a dataset: it has a protocol, a codebook, a reliability figure and a limitations section, and it makes a claim rather than publishing a table.
- State the exclusion funnel with counts. Silent filtering is the most common defect in opinion research.
- Rates, never raw counts, for anything over time. The corpus grew; a raw count measures the corpus.
- Findings and hypotheses stay visibly separate.
- Say what would make this wrong, in every study, naming selection effects and corpus coverage specifically.
- Single-coded studies are exploratory and must be labelled as such.
- Never pool measures that point in opposite directions. A holding that language is ambiguous and a holding that it is clear are different findings, and so are a grant and a denial of reformation.
CITATION.cff carries the machine-readable form and GitHub renders it as a "Cite this repository" button.
Klagge, Kevin D. legal-empirics: measuring which drafting language gets litigated. https://doi.org/10.5281/zenodo.22247377
That is the concept DOI and always resolves to the newest version. Cite it in prose. Use the version DOI printed on a specific Zenodo record when pinning a number to the exact release it came from.
The datasets are cited separately and will carry their own DOIs. The DOI above identifies the code and the methodology, which is a different object.
CC BY 4.0 for text and derived data. Not legal advice; empirical research about published opinions creates no attorney-client relationship.