The Proposition-Usability Model

The framework

Verification as usually taught asks one question: is this case good law? The Proposition-Usability Model, proposed by Ross Brodskiy and Nathan Pokov in Good Law for What?, replaces it with four questions asked about each load-bearing proposition.

The starting point is a mismatch of units. Research tools index cases; lawyers cite propositions. A proposition is the operative object: a rule, a standard, an element, a burden allocation, an exception, an application of a factor to facts, or a refusal to extend a doctrine. One opinion holds many of them, and they can favor opposite parties, address different procedural stages, rest on different sources of law, and have different futures. A clean citator report is evidence that the case has not been overruled or reversed in a way the service captured. It is not evidence that the sentence about to be quoted states current law, in this forum, for this client.

The four questions

1. Is the proposition still good law, as distinct from the case? A later court may overrule one holding, distinguish a second and follow a third. For a clean case, ask whether any later decision narrowed, limited, or added prerequisites to this point. For a flagged case, ask whether the negative event touched this holding at all. The hard residue is what the paper calls the Kisor class. Kisor v. Wilkie declined to overrule Auer deference but conditioned it on a series of gates, and a doctrine that had long worked mostly for the government began offering much of its best material to challengers. Nothing was overruled, so no treatment event recorded the change.
2. Does it govern in this forum, at this procedural stage? Forum and posture travel together, because the worst failures happen where two bodies of law describe the same procedural moment in the same words. The tell is a persuasive citation from another system quietly doing the work of a controlling one.
3. Whom does it help? In one sentence: if opposing counsel found this proposition, what would they do with it? Legal materials are adversarially structured. A holding that the employer bears the burden on an affirmative defense is not a neutral description; it is an allocation, and it favors the employee. The AI and law tradition built this in decades ago (Ashley's HYPO, Aleven's CATO), yet no major platform treats direction as a queryable attribute of an individual point of law.
4. Has this argument already been rejected here? A separate search, framed around the contention rather than the support, for courts in the jurisdiction that refused it. A rejected contention is rarely a headnote and often triggers no citator signal. This is the profession's negative space, and no product prompts the researcher to look for it.

Every answer rests on a citation or a recorded search (source, query terms, date), so that "nothing found" is a documented result rather than an assumption.

What the model claims, and what it does not

None of the four questions is new. Headnote-restricted citator reports, jurisdiction and motion-type filters, ordinary case analysis, and opposition research each cover part of the ground. The claim is the assembly: one named protocol, teachable in a single class, stated without reference to any system's architecture, and applicable to research whose first draft a machine wrote. The paper also argues that the gap is representational, not editorial. Even a flawless citator records treatment events, and three of the four questions are not about treatment at all.

For judging research tools, the paper adds a distinction. A tool is structurally reliable for a class of error when the information needed to catch it is recorded and a defined, repeatable check consults it. It is opportunistically reliable when catching the error depends on what a given query happens to retrieve. Loper Bright is a weak test, because the overruling of Chevron saturates recent text; Kisor is the strong one. The paper turns this into seven demonstration questions for vendors: compound constraints honored before results appear, validity tracked below the case, doctrines with shared vocabulary kept apart, rejected arguments searchable as objects, an inspectable reasoning trace, disclosed coverage and update latency, and human editorial provenance.

How to apply it

The paper's own classroom exercise is the best worked example. The instructor states a representation, "We represent the defendant moving for summary judgment in the Superior Court of California," and has a generative tool draft a one-page memorandum on the movant's initial burden. Every citation is confirmed real before class; the exercise is about usability, not fabrication. Suppose the memo supports the burden with a federal decision stating the Celotex formulation: a moving defendant may carry its initial burden by showing an absence of evidence supporting the plaintiff's case. Students first list the memo's propositions (not its cases), then complete a table for each.

The deliverable adds a paragraph per proposition, either defending the memo's use of the authority or rewriting the sentence so it becomes usable; here, restating the burden under Aguilar. The paper predicts the pattern instructors will see: strong answers on question one, uneven answers on question two, and initial confusion on question four, the least natural of the four and the reason the exercise exists. One variation assigns half the class to the movant and half to the opposing party. They will answer question three differently about the same propositions, which is the point.

Known limitations and critiques

The model names questions; it does not answer them. Each question costs reading time, and a brief with thirty load-bearing propositions multiplies that cost. The model makes skipped work visible. It does not make the work cheaper.

The Kisor class stays hard. Question one reaches changes in a proposition's practical force that no treatment vocabulary records. Naming the question does not supply the data: a researcher either knows the shift or has to find commentary describing it.

Direction is not always one-dimensional. The one-sentence test works for a clean burden allocation. A proposition that helps on one element and hurts on another, or helps now and constrains later, resists a single answer. The rubric rewards the sentence; practice sometimes needs a paragraph.

A negative search cannot prove absence. A recorded search that finds nothing is documented, not conclusive. Mart's work, cited in the paper, found that identical searches on six databases returned top-ten results with little overlap, and only about seven percent of cases appeared in all six. Question four depends on the product and on the drafter's guess at wording.

Posture is folded into forum. Question two treats forum and posture together, which fits the Celotex and Aguilar example. Mismatches of standard, record, and burden across procedural stages deserve their own treatment; see the posture-mismatch taxonomy.

Existence checks still come first. The exercise removes fabricated citations on purpose. In practice the four questions sit on top of confirming that each case exists and says what the brief claims.

Conflict of interest. The lead author discloses a financial interest in software that performs proposition-level extraction and citation validation. The paper's response is to state a standard any librarian can apply without vendor cooperation, and one the author's own products can fail. Readers should still weigh the disclosure.