Propositions, not documents. Verified once, at ingestion.
This is the one page on the site that leads with mechanism: how proposition-level legal AI is built and verified. Everything else starts from the case you almost lost.
The atomic unit of law is the proposition.
An attorney does not cite a case. She cites a proposition and offers a case as authority for it. Document-level tools store the case. Solon stores the proposition.
Retrieval systems built for document question-answering chop a forty-page opinion into chunks and hand the model whatever is nearest. A chunk cannot know whether a sentence is the holding or an aside, whether it favors plaintiff or defendant, whether the court established the rule or merely applied it, or whether it was narrowed last year. Legal reasoning operates on propositions, not topics.
Verification moved from query time to ingestion time.
Hallucination risk is moved from query time, where it is invisible, to ingestion time, where it is auditable and fixed once.
Solon read the published corpus once and pulled out what each case actually holds, whom it favors, and whether it is still good law, locking every proposition to the exact source lines it came from. When a lawyer drafts, they build on statements already verified, not on a model’s memory. A correction propagates to every future query. The residual is one-time extraction error, which we measure and publish rather than hide.
The published validation and its figures are in The Promise Fulfilled (SSRN).
Typed, adversarial metadata on every proposition.
Queries of the form “authority favoring this party on this point, at this posture, in this forum, still good law” are answerable because every filter is a stored field, not a guess at query time. When nothing satisfies every constraint, the honest answer is nothing.
Authority status
Reversed, modified, vacated, review granted, abrogated by statute, overruled, narrowed, and the qualified middle between fully valid and fully dead.
Litigation stage
Pleading, discovery, summary judgment, trial, appeal. A summary-judgment holding never masquerades as authority on a motion to dismiss.
Party favorability
Which side a holding helps, calibrated to future litigation rather than to who won the source case.
Burden
Who bears it, what triggers a shift, what shifts, and what stays put.
Rejected arguments
What the court declined to do, indexed as a first-class object.
Jurisdiction hierarchy
Binding versus persuasive in your court, the highest controlling authority on the point, department splits with the controlling line identified.
Doctrinal boundaries
Where a rule stops: distinguishing patterns, exceptions, the facts that took a case outside it.
Standard of review, materiality, and more
And more, described in the papers at the level that is safe to describe.
Verified against the source, with the source attached.
Every proposition in a memo is checked against the exact source lines it came from before it is ever used, and the source stays attached.
A validity check runs on every authority for subsequent history: reversed, modified, vacated, review granted, abrogated by statute. A status problem stops the case before it can reach your draft, and the next currently citable authority takes its place. The memo carries a citation table you can hand to a judge if anyone ever asks how you knew.
We do not publish the internal categorization the system uses. What we publish is the result: the law you draft on has already been verified.
What travels with every proposition
- The exact source lines it was drawn from
- The court that established it and where it is binding
- The procedural stage it was decided at
- Whom it favors on this point
- Its current status under subsequent history
Your citator flags the case. We flag the sentence.
A case can stay green at the document level while the line you cited has been hollowed out.
A document-level citator flags a case. Solon flags the sentence. Authority health is a composite over explicit negative treatment, the density of limiting and distinguishing decisions, dissent signals and citation trajectory, computed for the proposition, not the opinion. The distinction between complete invalidity (a rule overruled) and qualified validity (a rule that now carries five prerequisites it did not carry before) is a stored fact, not a judgment call at midnight.
The clearest worked example is the narrowing of agency-deference doctrine: the case is still cited, the flag is still green, and the proposition litigators relied on no longer does the work it did.
Abrogation propagation. When a court expressly overrules a rule, Solon traces the later propositions that were standing on it, including the ones the overruling court never named and no citator flags. The default is “not affected” unless the shared rule can be stated in both courts’ own words, because calling good law dead is worse than missing a casualty.
The treatment classification standard →Four questions any verified proposition answers
- Is it still good law?
- Does it govern here, at this procedural posture?
- Whom does it help?
- Has it already been rejected in this jurisdiction?
From Good Law for What?, the proposition-usability model.
Standard search finds holdings. We find rejections.
Knowing what courts have refused to do is as important as knowing what they have done.
Rejected arguments are indexed as searchable objects with the argument type, the rejecting court, the reason and the scope. Before you commit to a theory, the system checks the negative space: same argument, same court, already lost. The adverse case arrives from us, with the distinguishing patterns to brief around it, and not from opposing counsel.
“I can’t find the answer” is not “the answer does not exist.”
The goal of legal AI is not to answer every question. It is to know which questions have no single answer.
There are two kinds of uncertainty. Shallow ambiguity means the controlling authority exists and has not been found yet; better retrieval solves it. Genuine ambiguity means the law itself is contested: circuits split, courts hedge, dissents are vigorous, doctrine drifts. Presenting a contested question as settled is worse than a fabricated citation, because every citation is real. A tool that erases contestedness erases the very property that determines how much independent judgment its answer requires.
Model confidence is not doctrinal certainty. A perfectly calibrated model can be highly confident about a question on which the circuits have split, because confidence measures the model. Whether the law is contested is a property of the law, so it is detected from what courts wrote, tied to specific opinions, never inferred from how sure a model sounds.
The MSJ Qualifier is built on the same rule. A summary judgment pathway that depends on unsettled law, or on evidence nobody has located yet, stays UNRESOLVED. UNKNOWN is never quietly turned into NO.
Five sources of genuine ambiguity
- Semantic
- The words admit more than one reading, and courts have construed the same term differently.
- Normative
- Competing principles of similar weight point to different outcomes, and courts say they are balancing.
- Methodological
- Different interpretive approaches reach different results from the same text.
- Jurisdictional
- Courts in different jurisdictions answer the identical question differently.
- Analogical
- Competing precedents support different outcomes, and the choice of analogy decides the case.
What the system does when the law is genuinely unsettled
- Detects the signals courts themselves leave: explicit split acknowledgment, judicial hedging, dissent vigor, dueling authorities.
- Aggregates across cases: stance inversion, jurisdictional divergence, temporal drift.
- Reports the question as unsettled and names the split, instead of silently picking a winner.
- Shifts strategy from dispositive argument to risk-assessed argument, and tells you so.
The framework is set out in Detecting Genuine Doctrinal Ambiguity. Its benchmark is promised, not yet released.
No language model writes a citation or a validity verdict.
What the model is allowed to do
Decompose a question, ideate, plan, synthesize prose from a registry of confirmed authorities. The swappable job. A smarter model upgrades it, and we are long that trend.
What the model is forbidden to do
Emit a citation. Decide validity. Both are deterministic lookups against a corpus we own. You cannot bake next month’s reversal into last month’s weights; the law is a mutable graph and weights are a photograph.