A 2025 preprint, "Controlling Authority Retrieval: A Missing Retrieval Objective for Authority-Governed Knowledge" (arXiv:2604.14488), formalizes a critical insight: the controlling legal document is often not the most semantically similar document. A later case overruling an earlier case may use nearly identical language to the original. Vector retrieval returns whichever has higher cosine similarity, with no awareness of the supersession relation between them.

The Stanford study documents a specific failure of this type: Westlaw's RAG system mischaracterized a 10th Circuit holding on a statutory subsection as a finding that the entire statute was unconstitutional, when the 10th Circuit had actually rejected that argument. The retrieved passage was genuinely relevant. The generated summary of the holding was wrong. Standard RAG optimizes for retrieval. It has no mechanism to verify the characterization of what was retrieved.

The paper calls this the "authority-ordering problem": the retrieved document must be the one that controls the question under a formal precedence or supersession relation, not the one that matches its surface text. Embedding cosine distance addresses only the second.

The fact that Loper Bright Enterprises v. Raimondo (2024) overruled Chevron U.S.A. v. Natural Resources Defense Council (1984) is not written anywhere in the original Chevron opinion. A system that retrieves and embeds source documents when they are first loaded has no way to reveal this relationship unless it has separately modeled the citation treatment graph.

"Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs" (arXiv:2606.00898, 2025) decomposes citation accuracy into three components: precision (does the provision exist?), relevance (contextually appropriate?), and temporality (valid at the relevant date?). Tested across five systems on 100 legal queries, 13 to 21% of generated citations were hallucinated. The temporal validity component was the most common failure mode, not outright fabrication. Hudson de Martim's "Beyond Probabilistic Similarity" (arXiv:2606.09724) names this "diachronic blindness": a retrieved regulation may be semantically relevant but have been repealed. RAG cannot know this from the document alone.

A federal circuit court opinion routinely runs 40 to 80 pages. The binding holding may occupy a single paragraph. The rest is factual background, procedural history, party arguments, analysis of prior cases, hypotheticals, and dicta. None of these carry precedential weight. At the document level, or at the chunk level, they are indistinguishable from holdings by surface text alone.

We measured this directly in our extraction runs across the 1951–1953 federal corpus. Less than 9% of sentence-level text classified as a holding. When a retrieval system returns the full opinion as context, it is handing the model roughly 91% noise by this measure (factual background, procedural history, arguments of counsel) alongside the one or two sentences that actually matter.

"LegalBench-RAG: A Benchmark for Retrieval-Augmented Generation in the Legal Domain" (Pipitone, Alami et al.; arXiv:2408.10343, 2024) found that standard reranking pipelines were inferior to not reranking after retrieval in legal contexts. Legal retrieval requires extracting minimally relevant, precisely typed text segments. Off-the-shelf retrieval optimization degrades this precision because it optimizes for broad semantic relevance, not for the narrow assertion that is legally operative. The document does not label its own parts. Without sentence-level type classification at extraction time, HOLDING, REASONING, DICTA, and PROCEDURAL text are indistinguishable to any retrieval system.

The purpose-built tools evaluated by Stanford, with full legal data infrastructure and attorney review, still hallucinated on 17 to 33% of straightforward queries. The Stanford authors explicitly note these figures are likely lower bounds. The failure is not retrieval quality. It is that the document does not contain what a verification system needs.

A 9th Circuit holding is binding precedent within the 9th Circuit. It is persuasive authority elsewhere. A 5th Circuit case on the same question may use more similar language to a query than the controlling 9th Circuit case, and thus rank higher in vector retrieval. The system has returned the wrong authority, not because its retrieval failed, but because authority is an ordering relation that embedding distance cannot represent.

The "Controlling Authority Retrieval" paper formalizes this: jurisdiction is an authority-ordering constraint. Metadata filtering on circuit can partially address it, but fails when out-of-circuit cases are genuinely more similar than in-circuit authority, which is precisely the case where filtering would matter most. A legal verification system must model jurisdictional scope as a first-class field, not as a document-level filter applied after retrieval. A 2025 EMNLP evaluation found a compounding problem: chunks from entirely wrong source documents are especially common in legal corpora because all judicial opinions share structural boilerplate; the very homogeneity that makes law internally consistent makes it adversarial for vector retrieval.

"HalluGraph: Auditable Hallucination Detection for Legal RAG Systems via Knowledge Graph Alignment" (arXiv:2512.01659, December 2024) documents a critical failure mode: BERTScore, the standard semantic similarity baseline, performs near chance (AUC 0.50 to 0.60) for legal hallucination detection. It is insensitive to entity substitutions. Swapping "plaintiff" for "defendant," "2024" for "2025," or "the court holds" for "the dissent argues" barely moves cosine distance. In legal text, these substitutions are fatal errors.

When knowledge graph alignment is used instead, treating relational structure rather than embedding proximity as the ground truth, AUC improves to 0.89. That 39-point gain does not come from better retrieval. It comes from treating legal text as a relational structure where party names, dates, and speaker attribution carry meaning, rather than as tokens whose proximity implies agreement.

Courts speak in multiple voices in a single document. The majority opinion, a concurrence, and a dissent all appear in the same file. Correct speaker attribution requires dependency parsing and coreference resolution at the sentence level, tracing the pronoun "we" to the judicial speaker through the preceding paragraphs. A passage three paragraphs into a dissenting section, where "we" refers to dissenting justices, is indistinguishable from a majority holding by surface text alone.