From search to litigation intelligence.

Research argument

The retrieval gap in litigation

Finding text and finding a legally useful comparison are different retrieval problems.

A conventional query begins with a known handle: a statute, citation, doctrine, party, phrase, or topic. A live matter often begins before those handles are settled. The legal team must identify which facts are material, what evidence changes the issue, which procedural posture controls, and which prior matters are comparable for the specific task at hand.

The legal case retrieval literature treats this as a distinct research problem, not a larger version of web search [10]. LePaRD frames passage retrieval as predicting passages from precedential decisions given the context of a legal argument; its strongest reported models reached 59 percent recall when trained over the 10,000 most-cited passages [2]. LegalSearchLM reports that lexical and embedding-based matching can produce legally irrelevant cases when legal elements are not represented [3]. For India, IL-PCSR is important because it models prior-case and statute retrieval jointly, reflecting their dependence in practice [6].

These findings motivate, but do not prove, the Jurius theory. The proposed mechanism is information preservation plus constraint enforcement. Decomposing a matter should retain signals that a single query string discards. Predeclared eligibility rules should then prevent a high surface match from outranking a lower surface match that is usable for the stated research task. The theory predicts gains only where those dimensions are both observable and relevant.

Problem statement: retrieve a small, reviewable set of cases that are similar for a stated legal purpose and valid within a stated authority context.

Research argument

Theory: constraint-aware matter similarity

A case is useful when issue-bearing similarity survives legal authority constraints.

Let q be the current matter, c a candidate case, and t the lawyer's research task. Jurius represents q and c across separate dimensions: facts, legal issues, evidence, procedure, court, authorities, reasoning, outcome, and later treatment. The dimensions are versioned and source-linked because facts develop and authority changes over time.

The theory separates graded similarity from eligibility. Dimension scores describe how the cases resemble one another for task t. The eligible set E_t contains cases satisfying the mandatory jurisdiction, court, time, posture, treatment, and corpus rules declared for that task. Inclusion means eligible for review, not legally controlling. A candidate can therefore be factually close yet excluded, or less obvious yet retained because it is eligible and shares the decisive issue.

This approach has an important intellectual predecessor. HYPO represented disputes through legally salient dimensions and compared a current fact situation with prior cases to expose relevant similarities and differences [12]. Jurius does not claim to invent dimensional case comparison. Its narrower contribution is a source-linked, temporally versioned retrieval design and an evaluation program for testing that design at corpus scale.

The causal claim is deliberately narrow: compared with a keyword query or one undifferentiated representation, a multi-view representation plus explicit eligibility rules should improve the quality of the review set. It does not claim that every dimension matters equally, that a graph always helps, or that similarity determines a legal outcome.

• E_t is the task-specific eligible corpus; it encodes research filters, not a conclusion that a case controls.
• Dimension scores s_d are normalized, weights w_d sum to one, and both are fixed before an evaluation run.
• The uncertainty term u(c) increases with missing sources, low extraction confidence, or unresolved treatment status.
• R is an ordering rule for retrieval experiments, not a probability of relevance, legal correctness, or outcome.

Research argument

From theory to retrieval design

The design combines signals because each one fails in a different way.

Candidate generation combines exact terms, semantic matches, legal elements, citations, entities, court hierarchy, and time. Exact terms preserve precision but miss paraphrases. Semantic matches recover paraphrases but can confuse legal and conversational similarity. Citation structure reveals lines of authority but may privilege frequently cited or older cases. The system should log the contribution and failure mode of every signal.

Authenticated corpus boundaries are part of the retrieval method. External sources can make knowledge updateable and traceable [1], but unrestricted retrieval may introduce noisy or non-authoritative material; CaseFacts reports this failure in a legal fact-verification benchmark [8]. Jurius therefore separates source eligibility from ranking and records the corpus searched for each result.

The output is an inspectable neighborhood. A lawyer can compare dimension profiles, open the complete record, and follow citation or entity links to bridge cases. Logic-rule work in legal case retrieval supports the importance of explanations tied to the actual ranking process [11]. Graph expansion remains an ablation to be tested, not an assumed improvement.

A valid explanation is a reproducible mapping from a similarity claim to ranking signals, source passages, and authority rules.

Research argument

Recursive research and collaborative case building

The most important research question often appears only after the first relevant case is opened.

Litigation research is recursive. A lawyer may begin with the whole matter, discover that one witness statement or procedural defect drove a historical outcome, and then search for cases where that sub-point was dominant. Any verified field or passage should be able to become a new query without erasing the path that produced it.

Useful findings should not disappear into a transient answer. A lawyer can record the proposition, its relevance, supporting and adverse passages, relation to the current matter, and review status in a shared case table. Other members of the team can challenge the rating, add contrary authority, or assign follow-up research.

This loop also produces evaluation evidence. Accepted and rejected candidates, reasons for rejection, research branches, and later corrections form a task-specific record of professional judgment. That record can test whether the representation improved research or merely added interface complexity.

• Re-search from a fact, item of evidence, person, authority, argument, or outcome.
• Preserve positive, negative, and neutral treatment rather than only favorable material.
• Record author, timestamp, source, relevance rationale, and review status for each finding.
• Export a defensible research trail without exporting private matter data by default.

Research argument

Source-bounded strategy testing

The defensible unit is a cited claim and its rebuttal, not a simulated verdict.

Once a case table has been reviewed, the team can organize claims, rebuttals, dependencies, and evidence gaps. That work may be performed by lawyers, software agents, or both. The method does not depend on a theatrical debate. It depends on a fixed source set, explicit roles, and a rule that every factual or legal proposition must point to an allowed passage.

Multi-agent debate improved reasoning on some general benchmark tasks [7], but that evidence does not transfer automatically to law. The relevant Jurius hypothesis is narrower: independent proposal and criticism may increase the number of supported counterarguments found at a fixed research budget without increasing unsupported claims. A single model, a checklist, or a second human reviewer may perform just as well and must be included as baselines.

Outcome prediction is outside the present claim. Any future estimate would require separate calibration by jurisdiction, matter type, time, and data availability. The immediate research output is a structured argument map with cited support, cited adverse material, confidence labels, and unanswered questions.

Success means more supported counterarguments and earlier discovery of evidence gaps at the same review budget. It does not mean a predicted winner.

Research argument

Trust, provenance, and professional control

A citation is useful only when the system preserves what the source says, who said it, and whether it remains valid.

A preregistered evaluation of leading legal research products reported hallucination rates between 17 and 33 percent despite retrieval-augmented designs [4]. A study of general-purpose language models answering verifiable questions about United States federal cases reported rates from 58 to 88 percent in its tested settings [5]. Neither study measures Jurius, but both reject the assumption that fluent legal text demonstrates accuracy.

The control model is provenance first. Users must be able to inspect the passage, document version, speaker or legal actor, extraction method, authority status, and transformation behind a proposition. High-risk actions require human confirmation. Matter data should be isolated by organization, encrypted in transit and at rest, access-controlled, and excluded from training unless a customer expressly agrees under a separate policy.

Operational governance should be continuous, consistent with the govern, map, measure, and manage structure of the NIST AI Risk Management Framework [9]. Deployment also requires jurisdiction-specific analysis of professional responsibility, privilege, confidentiality, data protection, document rights, and court rules.

• No uncited legal proposition in a research deliverable.
• Clear labels for allegation, evidence, party submission, judicial finding, and model inference.
• Human approval before a finding enters a final work product.
• Audit logs for retrieval, generation, edits, exports, and access.
• Routine red-team tests for fabricated citations, authority drift, prompt injection, leakage, bias, and overconfidence.

Research argument

A falsifiable evaluation program

Each stronger claim requires a stronger study; architecture alone earns no performance claim.

Evaluation begins with preregistered hypotheses, frozen test sets, written relevance criteria, and temporally held-out cases. Retrieval experiments compare BM25, a single dense representation, a multi-view representation, explicit eligibility rules, and graph expansion as separate ablations. Assessors should be blinded to system condition and include practicing lawyers or qualified legal researchers.

Primary outcomes are lawyer-rated relevance at a fixed review budget, recall of a gold case set, invalid-authority rate, and time to a defensible set. Results must be reported by jurisdiction, court level, matter type, language, document quality, and time. Test sets should contain adverse authority, overruled decisions, incomplete records, contradictory passages, and matters with no close precedent.

Before data collection, each study should publish its unit of analysis, primary outcome, smallest effect of interest, sample-size or power rationale, assignment method, exclusion policy, and analysis plan. Confidence intervals and assessor agreement should accompany point estimates. Gold case sets should be built through pooled independent research and adjudicated disagreement; absence from the pool must not be treated automatically as irrelevance.

Workflow and strategy tests come later. They should compare the system with ordinary search, a structured checklist, and a shared spreadsheet. Added complexity is justified only if it improves retrieval or decision usefulness at a comparable time and review cost.

Retrieval quality

Recall@k, nDCG@k, MRR, authority validity, and lawyer-rated relevance

Explanation fidelity

Whether stated similarity reasons reproduce the actual ranking signals and source passages

Research efficiency

Time to a defensible case set, duplicate work avoided, and useful findings per review hour

Generation safety

Citation precision, unsupported-claim rate, abstention quality, and temporal-validity errors

Argument coverage

Distinct supported theories, rebuttals, failure modes, and unresolved evidence gaps

Research argument

Boundary conditions, alternatives, and roadmap

Confidence comes from knowing where the theory should fail and what simpler explanation might win.

The theory should produce little or no advantage for simple citation lookup, narrow questions with stable vocabulary, or corpora whose metadata and full records are too poor to populate the required dimensions. Explicit structure may also slow research when the task is trivial. Graph expansion may amplify popularity bias. Eligibility rules may reduce recall if metadata is wrong. These are predicted failure modes, not implementation footnotes.

Competing explanations must be tested. Any gain may come from a larger corpus, more compute, a better interface, or additional review time rather than the matter representation itself. Ablations and matched-budget studies are therefore essential. The strongest alternative is that a well-tuned lexical-plus-dense ranker and a disciplined spreadsheet achieve the same lawyer outcomes with less complexity.

The credible roadmap is narrow. Begin with a defined Indian jurisdiction, court set, matter type, and lawfully usable source collection. Establish provenance and authority treatment, then test the representation and gates. Add recursive research and shared case tables only after retrieval gains hold. Evaluate source-bounded strategy testing last. Cross-jurisdiction expansion should follow measured differences in authority, language, procedure, and data quality.

Current claim level: Stage 0. Jurius specifies a theory, architecture, and evaluation program. It does not yet claim measured retrieval or workflow superiority.