Proposition Support Verification: How to Check What a Case Actually Stands For
Proposition support verification is the second-order check that comes after citation existence. There are two distinct failure modes in AI-assisted legal drafting. The first is the cite that does not exist. The reporter volume is named but does not contain the opinion. The parties are named but never appeared in that court. This is the failure mode of Mata v. Avianca, and it is the easier one to detect. A scanner that resolves the cite against reporter and public-record sources returns a clean negative verdict.
The second failure mode is the cite that exists, has the right parties, resides in the named reporter, and does not stand for the proposition the brief offers it for. The case is real. The citation resolves. The holding is somewhere else.
The second mode is harder. It is less discussed in bar opinions and standing orders, and it is arguably more dangerous, because it survives every existence check the profession is currently building. The brief reads correctly. The footnote resolves. The partner skimming the citation sees a real case. Opposing counsel opens the opinion and finds that the holding addresses a different question.
The two-layer model
A useful frame for the verification stack is two layers, sequenced.
The first layer is existence verification. Does the case exist? Does the reporter contain it? Do the parallel cites resolve? Did the named judge sit on the named panel? This is the layer Mata v. Avianca failed. A pre-filing scanner that resolves citations against reporter and public-record sources returns a verdict per citation. The output is binary on the underlying question of existence, even where the verdict language is hedged.
The second layer is proposition verification. Granted that the case exists, does the case stand for what the sentence in the brief claims it stands for? This is a different question, and the mechanism that answers it is different. Existence verification asks the reporter. Proposition verification asks the opinion.
Citation existence checking is becoming commoditized. Multiple tools now run reporter resolution. The harder problem lives at the second layer. That is where the next stretch of the verification surface needs to be built.
Why proposition verification is harder
Existence resolves to a record. A citation either matches a published opinion in the reporter or it does not. The verdict is a lookup against a structured source.
Proposition support is not a lookup. The verdict requires reading the opinion against the brief and answering several adjacent questions in sequence. What did the court actually decide? On what procedural posture? On what facts? What part of the language the brief quotes is the holding, what part is the court's recitation of an argument it then rejects, and what part is dicta? Has the case been narrowed, distinguished, or overruled by later authority?
Each of those is a different judgment. Consider an illustrative hypothetical. A brief cites a hypothetical Smith v. Jones for the proposition that a statute of limitations runs from the date of injury rather than discovery. The case exists. The court did discuss the limitations period. The language the brief quotes is a sentence from the court's recitation of the plaintiff's losing argument, not from the holding. The case in fact stands for the opposite of what the brief offers it for.
That error is not a hallucination in the strict sense. The model did not invent the case. The model retrieved the case and attached it to the wrong sentence. The existence check passes. The proposition check is the one that catches the error.
The same pattern recurs in less dramatic forms. A case is cited for a holding that is in fact dicta. A case is cited without acknowledging that subsequent treatment has narrowed it. A case is cited for a procedural-posture-specific holding and applied to a different posture. Each of these is a real risk, and none of them resolves at the existence layer.
What the adversarial review pass does
The mechanism that addresses proposition support is an adversarial review pass aligned with Model Rule 3.3(a)(2). After existence verification confirms a case exists, a second pass re-reads the cited opinion against the proposition the brief offers it for. The pass approaches the citation in the posture of an opposing-counsel reviewer asking whether the cite holds.
The pass produces several distinct observations per citation. It identifies the holding of the case as the court itself framed it. It notes the procedural posture, because the same legal question can resolve differently on a motion to dismiss than on summary judgment. It distinguishes the holding from dicta. It then records a verdict on whether the case, on those observations, supports the proposition the brief offers it for.
The pass does not confirm subsequent negative treatment. Whether a located opinion has since been overruled, vacated, or limited by later authority is a question a Shepardize or KeyCite check answers, and the attorney should still run that check before filing. Veritas reports what the opinion itself says about the proposition. It does not replace a citator.
The verdict is hedged in the same register as existence verdicts. The pass reports what it observed in the opinion. The pass does not overstate the verdict. The partner reads the result and applies legal judgment to the disposition.
Verdict states at the proposition layer
The verdict taxonomy at the proposition layer extends the taxonomy already in use at the existence layer. The full set is covered in the verification status taxonomy. One verdict is particularly relevant here.
Cited for unrelated proposition. The case exists. The citation resolves. The reviewer's reading of the opinion does not locate support for the proposition the brief offers it for. The verdict does not assert that no support exists anywhere in the opinion. The verdict reports that the pass did not find the support, and recommends that the partner re-read the case against the sentence.
The verdict is hedged because the underlying question is one of legal judgment, and the certificate is written to avoid asserting more than the pass can establish. One question the pass does not answer is subsequent treatment. Whether a located opinion has since been narrowed, distinguished, or overruled is a citator question that a Shepardize or KeyCite check answers, and the attorney should still run that check before filing.
What the partner sees on the certificate
The Verification Certificate with proposition-level verdicts reads as a structured list. Each row corresponds to a cited proposition rather than a raw citation. The columns include the cite, the sentence in the brief that the cite supports, the existence verdict, the proposition verdict, the recommendation, and the source records consulted.
The partner reading the certificate is not asked to re-read every cited opinion. The partner is asked to attend to the rows where the verdict is something other than clean support. A row that reads cited for unrelated proposition is the row the partner opens the case behind, re-reading the opinion against the sentence before deciding whether the cite belongs.
The job is structured. The unstructured review of forty footnotes becomes a focused review of the three rows that the adversarial pass flagged. The partner retains every judgment call about whether the cite belongs in the brief. The certificate narrows where the judgment needs to be applied.
The next frontier
The first wave of AI-citation tooling has settled on the existence problem. That work is real, and Mata v. Avianca is the case that made it unavoidable. The second wave is the proposition problem. It is harder to mechanize, harder to render in clean verdicts, and harder to audit. It is also where the more interesting category of error now lives.
A complete pre-filing verification surface runs both layers. The existence pass returns a verdict on whether the case can be located. The adversarial review pass returns a verdict on whether the case stands for the proposition. The certificate records both. The matter file carries the record.