The field wrote the math. We built the instrument.
Retrieving structured legal propositions instead of documents is a nearly seventy-year-old idea the industry detoured around, because chunk-embed-retrieve ships a demo by Thursday and the law’s requirements don’t.
Fifty years of theory. Zero production systems. Until now.
The first proposal that legal text could be normalized into a form a machine could reason over.
Rules as computable structures, decades before anyone could extract them at scale.
Case-based reasoning that knows which side a factor helps. The requirement never went away.
Defeasible argumentation: rules with exceptions, burdens and priorities, formalized.
What a precedent forces on a later court, stated precisely. The specification was complete.
Frontier models made proposition-level extraction across millions of opinions economical. The bottleneck moved from formalization to trust, integration and scale.
The gap was always infrastructural, not intellectual.
Thirty-five years of scholarship specified what legal reasoning requires: adversarial directionality, typed rules, structured precedential constraint. Document-level retrieval satisfies none of them.
The resulting failures are categorical, not incremental: authority-health blindness, jurisdictional conflation, doctrinal boundary collapse, directional opacity and negative-space ignorance. Agentic augmentation converts some of these from systematic to opportunistic and leaves the rest intact, because they are knowledge-representation failures, not coverage failures. Structural reliability, where the stored representation makes an error class impossible or reliably detectable, is different in kind from opportunistic reliability, where correctness depends on whether retrieval happens to surface the missing context.
Five structural failure categories
- Authority-health blindness
- Jurisdictional conflation
- Doctrinal boundary collapse
- Directional opacity
- Negative-space ignorance
The five categories are the conceptual backbone of the benchmark.
The ideas, without the footnotes.
Is legal AI just retrieval-augmented generation?
Mostly, yes, and that is the problem. Retrieval answers “what cases exist on this topic.” Reasoning answers “what is the law, who said it first, does it help me.” A chunk cannot carry holding versus dicta, stance, origin, category or treatment. A proposition can.
How Solon works →Why summary judgment is a harder problem than legal research
Research maps facts to propositions, which are system-generated and gradient. A motion maps facts to elements, which are fixed by law and binary. Summary judgment is a constraint-satisfaction problem under adversarial inference, not a similarity search.
The MSJ Qualifier →Two kinds of uncertainty
“I can’t find the answer” and “the answer does not exist” are different states, and legal AI that cannot tell them apart presents contested law as settled. Genuine ambiguity has five sources: semantic, normative, methodological, jurisdictional and analogical. It shows up in three layers: the signals a court leaves in a single opinion, the patterns that appear only when cases are aggregated, and the structure of the doctrine itself. The fix is architectural: detect and map indeterminacy rather than smooth it over.
Read the essay →AI as measurement infrastructure, not judge
The same technology that would be illegitimate as a judge may be legitimate as a measuring device: hold the verified law fixed, vary only the posture, and report how open a question is as an artifact for human review.
Ariadne, under review →Can I still rely on this case, and to what extent?
A flag on an opinion cannot answer that, because a litigator cites a holding, not an opinion. Treatment classification works at the level of the holding, along three dimensions: scope, severity and mechanism. Overruled is not abrogated, limited is not distinguished, and reversed is not a synonym for dead.
Treatment classification →When a “therefore” is law
Authority does not transfer from a decided case to a new one by itself. It travels by a small set of recognized moves, each with conditions. If a conclusion crosses none of them, it is argument, not law.
The twelve bridges →Twelve bridges from precedent to a conclusion.
Legal reasoning is constrained, not creative. A lawyer’s “therefore” cannot float free from precedent. It must rest on a recognized doctrinal bridge, or it is advocacy, not law.
When a brief says that two cases both protect some interest and therefore together protect a third thing, the court asks which bridge connects them. If none does, the conclusion belongs in the argument section, not in the statement of governing law. Twelve bridges cover the ground: rules applied directly, reasoning extended by analogy or principle, and the settled weight of stare decisis.
Moves no bridge will carry
- Cases grouped by theme with no shared holding.
- Dicta presented as the holding.
- A later gloss read backward into an earlier opinion.
- A dissent quoted as though it were law.
- A test lifted from one doctrinal area into an unrelated one.
- An implicit overruling asserted where the later court never said so.
This is the audit Opposition runs on every citation in a motion: does the authority satisfy the conditions of the move claimed for it?
Direct holding application
The facts sit inside the rule the prior case announced. Quote the holding, not the background.
Established multi-case doctrine
A controlling court already distilled the line into one test. Cite the synthesizing case, not your own synthesis.
Recognized analogical extension
The rationale, not the surface facts, carries over, and the analogy is made explicit.
Intra-doctrinal evolution
A line of cases inside one doctrinal family, each built on the last.
Controlling superior court adoption
A higher court adopted the rule. Cite the adoption, not the origin.
Statutory codification
The legislature wrote the judge-made rule into text. Cite the statute.
A fortiori
If the rule held in the weaker case, it holds with stronger reason in this one.
Negative inference
An exhaustive list excludes what it omits. Expressio unius, and only when the list is truly exhaustive.
Structural inference
The conclusion follows from the constitutional structure itself, as courts have already read it.
Overruling chain completion
A case was overruled in part. Identify exactly which holdings survive, in the later court’s own words.
Identified policy rationale
The court stated the purpose behind its rule, and that stated purpose reaches these facts. Only reasons the court itself gave.
Stare decisis weight
Reaffirmation and reliance have entrenched the rule. The argument is about settled weight, not a new rule.
Read the papers.
Five on SSRN, one of them under review. Link out, no PDFs hosted, abstracts here.