The problem a ranked list cannot solve
A legal research question is not a bag of keywords. "Does a non-compete survive a change of control where the employee never signed the assignment?" decomposes into doctrines, jurisdictions, procedural postures, and authorities of unequal weight. A binding decision from the controlling appellate court is worth more than a persuasive opinion from a sister jurisdiction, which is worth more than an unpublished disposition, which is worth more than a trial-court order that was later reversed. Keyword search and embedding search collapse this hierarchy into a single ranked list. The list is useful for browsing and unsafe for a brief, because the property a lawyer needs is not "looks similar to my query" but "is admissible authority for my argument in my forum, and I can show how I found it."
Three failures recur. First, relevance without standing: a returned case is on point doctrinally but has been overruled, depublished, or is from a jurisdiction whose decisions do not bind the forum. Second, results without a path: the lawyer cannot reconstruct why a particular case surfaced, which makes the research neither auditable for a malpractice record nor reproducible by a colleague. Third, fabricated authority: a generative system asked to summarize or extend research emits a citation to a case that does not exist, and nothing in the retrieval pipeline structurally prevents the fabrication from reaching the filing.
These are governance failures, and they are the failures the disclosed technology is built to remove.
What semantic discovery contributes
Semantic Discovery, disclosed in United States Patent Application 19/647,395, models the research process as a typed discovery object that traverses an adaptive index of authorities. The object is not a query string; it is a persistent structure with typed fields including intent, context, accumulated memory, governing policy, lineage, and confidence. At each anchor boundary in the index, the object undergoes a single three-in-one traversal step that fuses three operations the conventional pipeline keeps separate:
- Search identifies the candidate authorities reachable from the current anchor.
- Inference evaluates those candidates against the discovery object's intent and accumulated context, so the next move depends on what the research has established so far rather than on the original query alone.
- Governance is a constituent phase of the same step, not a post-filter bolted on afterward. The model proposes a transition; the substrate decides. Each proposed step is evaluated against policy and resolves to one of three outcomes: admit the transition, reject it, or decompose it into constituent sub-questions that are themselves traversed. Because governance runs inside the step, an inadmissible move never executes, rather than executing and being flagged later.
Every admitted step is written to the discovery object's lineage field, and the lineage is admissibility-verified: the complete traversal history is reconstructible from the field alone, and each recorded step carries the admissibility determination that permitted it. Traversal is query-specific and post-PageRank: the relevant neighborhood is computed for this question at this anchor, not read off a precomputed global popularity ranking that has no notion of jurisdiction or doctrine. Alias resolution lets the same authority referenced under different citation forms, reporters, or party-name spellings resolve to one node, so the research does not silently treat two names for one case as two cases. Per-step overhead is bounded, so governance is a structural property of traversal rather than a tax that grows without limit as the research deepens.
The application also draws on two inventions incorporated by reference in 19/647,395: the adaptive index that serves as the traversal substrate, which organizes authorities into entropy-banded anchor clusters with slope-validated lookup, and the policy-governance framework that supplies the signed, policy-bounded admission criteria the governance phase enforces.
How it maps onto legal research
The research question as a persistent discovery object. The matter, the doctrines in play, the controlling forum, and the authorities already relied upon are carried in the object's intent, context, and memory fields. Research conducted over days or across a team accumulates into one object rather than restarting from a blank query box each session. When a brief is later challenged, the object is the record of how the position was researched.
Jurisdictional and authority scoping as governance policy. "Binding in the Ninth Circuit," "good law as of the filing date," "published dispositions only," and "exclude authorities that have received negative subsequent treatment" are expressed as policy bounds the governance phase enforces at admission. A transition into an out-of-scope or overruled authority is rejected at the step, so out-of-scope cases do not enter the working set in the first place. Different matters and different forums carry different policies against the same corpus.
Decomposition for compound questions. A compound legal question that cannot be admitted as a single traversal is decomposed into its doctrinal and procedural sub-questions, each traversed under the same governance, and recombined. The decomposition itself is recorded in lineage, so the structure of the analysis is part of the auditable record.
Citation verification as a structural property. Because a returned authority is reachable only as the endpoint of admitted, lineage-recorded steps through a real index of real authorities, a citation that no authority supports has no admissible path and cannot be produced. The defense against fabricated citations is the traversal discipline, not a separate fact-checking pass run after generation. Alias resolution ensures that verification operates on the authority itself rather than on an unreconciled surface form of its name.
Embodiments and deployment options
The mechanism admits many faithful implementations, and the following are illustrative rather than exhaustive.
- Corpus scope. Federal and state case law, administrative decisions, agency guidance, regulations, secondary authorities such as treatises and restatements, a firm's internal work-product and brief bank, or a closed discovery production. The index is the traversal substrate; the corpus loaded into it is a deployment choice.
- Governance posture. Policies may be set per jurisdiction, per matter, per practice group, or per client, and may encode ethical-wall constraints so that a discovery object traversing one matter cannot reach authorities or work-product walled off from it. Negative-treatment and good-law-as-of-date constraints are policy parameters, not separate products.
- Mode of use. A lawyer-driven interactive session in which each admitted step is inspectable; an autonomous research agent that traverses under fixed policy and returns a lineage report; or a verification-only pass in which an existing draft's citations are each re-derived as admissible traversal paths and any citation lacking one is flagged.
- Integration surface. A standalone research interface, a drafting-tool plug-in that verifies citations as they are inserted, or a governance layer placed in front of a generative model so that the model's proposed authorities must clear admissible traversal before they reach the document.
- Output artifacts. A ranked-but-scoped working set, a memorandum with every proposition tied to its lineage path, or a machine-readable traversal record suitable for the documentation expectations of regulated AI use and for a malpractice or sanctions record.
- Multi-researcher coordination. Two discovery objects pursuing related questions on the same matter can be merged under policy, with conflicts resolved and the merge recorded in lineage, so a team's parallel research consolidates without losing provenance.
A skilled implementer can build any of these from the disclosed primitive: load authorities into the adaptive index, instantiate the research question as a typed discovery object, express jurisdiction and good-law constraints as governance policy, run the three-in-one traversal with admit, reject, and decompose at each anchor boundary, and read the admissibility-verified lineage out as the citation record.
Why this is the defensible architecture
The regulatory instruments named above do not ask for better search; they ask for governed, traceable, verifiable research. A pipeline that retrieves first and audits later cannot offer those properties as guarantees, because the inadmissible result has already been produced before the audit runs. Semantic discovery inverts the order: governance is a phase of the traversal step, the model proposes and the substrate decides, and the lineage that documents the research is the same structure that produced it. Relevance without standing, results without a path, and fabricated authority are each removed at the point where a step is admitted, rather than detected after a filing is at risk.
Disclosure Scope
This disclosure covers the application of Semantic Discovery, disclosed in United States Patent Application 19/647,395, to legal case research and related legal information work. Within scope are: the representation of a legal research question as a persistent discovery object with typed intent, context, memory, policy, lineage, and confidence fields; the three-in-one traversal step in which search, inference, and governance fuse at each anchor boundary with admit, reject, and decompose outcomes; the expression of jurisdictional, good-law, published-only, ethical-wall, and negative-treatment constraints as governance policy enforced at admission; query-specific post-PageRank traversal over an adaptive index of authorities; alias resolution across citation forms, reporters, and party-name spellings; admissibility-verified traversal lineage used for citation verification and as an auditable research record; decomposition of compound legal questions and multi-discovery merge under policy; and the embodiments and deployment options enumerated above, including interactive, autonomous, and verification-only modes and the integration of the primitive as a governance layer in front of a generative model. Implementations are within scope across federal, state, administrative, regulatory, secondary, internal work-product, and closed-production corpora. Pipelines that retrieve candidate authorities first and apply governance only as a post-retrieval filter, rather than as a constituent phase of the traversal step, fall outside the disclosed mechanism.
