SCALE: Structured Case-law-guided Anticipatory Legal Reasoning¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/khatrimann/SCALE
Area: LLM Reasoning
Keywords: Legal Reasoning, Retrieval-Augmented Generation, Case-Law Hierarchy, Anticipatory Retrieval, European Court of Human Rights
TL;DR¶
Addressing standard legal RAG's vulnerability to surface-level lexical similarity and the burden of synthesizing dispersed case judgments, SCALE compiles article-specific Case Law Guides into an Issue–Factor Hierarchy (IFH) to drive anticipatory, factor-specific precedent retrieval, markedly boosting judicial reasoning consistency and precedent faithfulness.
Background & Motivation¶
Large language models have shown growing potential across legal natural language processing, from legal judgment prediction to summarization. However, generating precedent-grounded legal reasoning that is factually faithful, legally sound, and consistent with judicial doctrine remains an open challenge. In courts interpreting fundamental rights such as the European Court of Human Rights (ECHR), legal adjudication demands far more than surface pattern matching: a judge must identify the governing legal issues under an alleged provision, isolate the legally salient factual factors, and weigh them against established precedent. Implicit, one-step generation by LLMs regularly suffers from hallucinations, unsupported factual abstractions, and misapplied legal authorities.
Standard legal retrieval-augmented generation (RAG) is commonly deployed to ground model outputs in external authority, yet it encounters two fundamental obstacles. First, raw retrieved case law is notoriously long, heterogeneous, and unorganized relative to the specific argumentative demands of the pending case, forcing the downstream model to filter and synthesize noisy context during generation. Second, retrieval is typically driven by surface similarity between the case facts and prior texts, even though true legal relevance hinges on latent legal questions and recurring balancing factors rather than verbatim factual overlap.
While recent anticipatory retrieval approaches attempt to predict legal principles prior to querying, unconstrained free-form abstractions remain brittle, especially for smaller models navigating nuanced proportionality doctrines. Core idea: externalize article-specific legal doctrine from Case Law Guides into an explicit Issue–Factor Hierarchy (IFH), using this structured scaffold to steer anticipatory, issue-aligned, and factor-specific retrieval and generation, thereby compensating for the weaker latent reasoning capacity of open-weight models.
Method¶
Overall Architecture¶
SCALE structures precedent-grounded legal reasoning within the ECHR jurisdiction. Instead of performing naive dense retrieval across tens of thousands of scattered historical judgments, SCALE harnesses official ECHR Case Law Guides, which synthesize decades of jurisprudence into core legal questions, relevant factual dimensions, and authoritative paragraph citations. The architecture executes in three distinct stages: offline Issue–Factor Hierarchy construction, online two-step issue and factor identification, and structured precedent-grounded reasoning generation.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
subgraph S1["Offline Doctrinal Structuring"]
direction TB
A["Case Law Guide Text Chunks<br/>(~2,000 tokens/chunk)"] --> B["Issue-Factor Hierarchy Construction<br/>Local IFH Extraction & Iterative Merging"]
end
subgraph S2["Online Identification & Targeted Retrieval"]
direction TB
C["Case Facts + Alleged Article"] --> D["Issue & Factor Identification<br/>Article Issue Filtering → Factual Factor Matching"]
B -.->|Structured IFH Index| D
D --> E["Linked Passage Retrieval<br/>Extract Verified Guide Paragraphs"]
end
subgraph S3["Judicial Opinion Generation"]
direction TB
E --> F["Reasoning Generation<br/>Legal Analysis + Proportionality + Conclusion"]
C -.-> F
F --> G["Final LAW Section Reasoning Trace"]
end
Key Designs¶
1. Issue–Factor Hierarchy Construction: Transforming Doctrinal Guides into Traceable Graphs
To resolve the challenge of synthesizing unstructured, sprawling case texts, this design converts article-specific Case Law Guides into an explicit hierarchical representation. Top-level nodes represent "Legal Issues"—the core normative questions that structure the Court's judicial review under a Convention article (e.g., whether state interference pursued a legitimate aim). Nested beneath each issue are "Legal Factors"—recurring factual considerations relevant to resolving that question (e.g., applicant vulnerability, severity of state action). Operationally, the guide is split into chunks of approximately 2,000 tokens. An LLM independently extracts local issue-factor hierarchies along with their supporting guide paragraph numbers (e.g., ¶9, ¶13). An iterative merge prompt then consolidates these local trees into a single unified article-level IFH, resolving synonyms and reconciling factor granularity while strictly preserving every paragraph reference link.
2. Issue and Factor Identification: Two-Stage Anticipatory Constrained Retrieval
To overcome the pitfalls of raw factual similarity search, this design decomposes retrieval into a top-down, two-stage identification process over the prebuilt IFH. In the first step (issue selection), given the case facts and the alleged Convention article, the model selects which overarching legal issues from the IFH are engaged. In the second step (factor selection), for each selected issue, the model evaluates its attached list of factors against the case facts, outputting only those factual factors directly corroborated by the record. Because every factor in the IFH is explicitly linked to authoritative guide paragraph citations, the system retrieves only the corresponding paragraphs from the source guide. This achieves anticipatory, issue-aligned, and factor-specific precedent retrieval without relying on unconstrained query expansion or noisy vector similarity over full case corpora.
3. Reasoning Generation: Precedent-Grounded Judicial Deliberation
To prevent superficial citations and unprincipled conclusions, the final generation prompt conditions the LLM to adopt the formal role of an ECHR judge constructing the official "LAW" section of a judgment. Receiving the case facts, party submissions, and the targeted guide passages retrieved via the IFH, the model is guided through a structured argumentative progression: defining the scope of the right and state obligations under the alleged article, measuring the factual record against the precise balancing tests established in the retrieved precedent (e.g., margin of appreciation, procedural safeguards, proportionality), explicitly articulating whether past authorities support or constrain the claim, and stating a formal ruling on violation. This eliminates ungrounded speculation by directly tethering the narrative to controlling authority.
A Worked Example¶
Consider an Article 11 (freedom of assembly and association) case concerning the refusal of state authorities to register an association (Case 243777): - Input: Domestic courts rejected registration of an association on the grounds that its chosen name contained an ethnic reference ("Turkish Women") that could mislead the public regarding members' nationality; - Issue Identification: SCALE selects the governing legal issue: whether the refusal constituted an interference prescribed by law, in pursuit of a legitimate aim, and necessary in a democratic society; - Factor Identification: From the IFH, SCALE isolates specific factual factors: the association's charter explicitly restricted membership to adult Greek nationals of that prefecture, the name reflected cultural identity rather than a threat to territorial integrity, and the domestic authorities failed to provide relevant and sufficient empirical evidence of public confusion; - Reasoning Generation: Unlike baselines that jump from domestic refusal directly to an abstract finding of unfairness, SCALE executes a structured proportionality analysis. Grounded in the retrieved guide paragraphs on association naming and state margins of appreciation, the model explains that domestic authorities failed to satisfy the proportionality requirement, applying the exact doctrinal standards of the Court to reach a robust finding of a violation.
Key Experimental Results¶
Main Results¶
Evaluation was conducted by an ECHR legal expert (a former Council of Europe Programme Advisor and case lawyer) on 37 case–article instances spanning 4 model configurations (Llama-3.1 8B/70B, Qwen3 8B/14B) and 6 methods, reviewing 888 reasoning traces on a 1–5 Likert scale across six criteria: Legal Issue Identification, Factual Faithfulness, Logical Coherence, Precedent Relevance, Precedent Application, and Judicial Alignment.
| Model | Method | Legal Issue ID | Factual Faithfulness | Logical Coherence | Precedent Relevance | Precedent Application | Judicial Alignment |
|---|---|---|---|---|---|---|---|
| Llama-3.1 8B | Parametric (PR) | 3.54 | 3.19 | 3.14 | 3.00 | 2.96 | 3.12 |
| CaseRAG | 3.46 | 3.28 | 3.23 | 3.24 | 3.19 | 3.26 | |
| A-CaseRAG | 3.72 | 3.42 | 3.34 | 3.50 | 3.43 | 3.39 | |
| GuideRAG | 3.88 | 3.39 | 3.31 | 3.66 | 3.58 | 3.47 | |
| A-GuideRAG | 4.05 | 3.57 | 3.43 | 3.87 | 3.81 | 3.64 | |
| SCALE (Ours) | 4.43 | 3.91 | 3.69 | 4.25 | 4.19 | 3.93 | |
| Llama-3.1 70B | Parametric (PR) | 3.70 | 3.33 | 3.29 | 3.16 | 3.14 | 3.26 |
| A-GuideRAG | 4.15 | 3.61 | 3.54 | 3.99 | 3.93 | 3.76 | |
| SCALE (Ours) | 4.39 | 3.98 | 3.82 | 4.29 | 4.23 | 4.03 | |
| Qwen3 8B | Parametric (PR) | 3.67 | 3.41 | 3.26 | 3.14 | 3.08 | 3.28 |
| A-GuideRAG | 4.22 | 3.74 | 3.48 | 4.06 | 3.99 | 3.80 | |
| SCALE (Ours) | 4.56 | 4.15 | 3.90 | 4.41 | 4.36 | 4.18 | |
| Qwen3 14B | Parametric (PR) | 3.80 | 3.53 | 3.41 | 3.26 | 3.21 | 3.40 |
| CaseRAG | 3.72 | 3.58 | 3.47 | 3.45 | 3.42 | 3.48 | |
| A-CaseRAG | 4.00 | 3.69 | 3.56 | 3.78 | 3.74 | 3.66 | |
| GuideRAG | 4.15 | 3.74 | 3.54 | 3.93 | 3.88 | 3.78 | |
| A-GuideRAG | 4.28 | 3.82 | 3.65 | 4.13 | 4.08 | 3.91 | |
| SCALE (Ours) | 4.59 | 4.24 | 4.00 | 4.47 | 4.42 | 4.27 |
Ablation Study¶
The component-wise progression isolates the impact of external knowledge sources, query abstraction, and explicit structure (evaluated on Qwen3 14B mean scores):
| Method Stage | Core Mechanism Shift | Legal Issue ID | Precedent Relevance | Precedent Application | Judicial Alignment |
|---|---|---|---|---|---|
| ① Parametric (PR) | Closed-book parametric inference | 3.80 | 3.26 | 3.21 | 3.40 |
| ② CaseRAG | Raw-fact queries over 17k historical cases | 3.72 | 3.45 | 3.42 | 3.48 |
| ③ A-CaseRAG | Principle-abstracted queries over historical cases | 4.00 (+0.28) | 3.78 (+0.33) | 3.74 (+0.32) | 3.66 (+0.18) |
| ④ GuideRAG | Raw-fact queries over Case Law Guides | 4.15 (+0.15) | 3.93 (+0.15) | 3.88 (+0.14) | 3.78 (+0.12) |
| ⑤ A-GuideRAG | Free-form anticipatory queries over Guides | 4.28 (+0.13) | 4.13 (+0.20) | 4.08 (+0.20) | 3.91 (+0.13) |
| ⑥ SCALE | Explicit IFH constrained issue/factor retrieval | 4.59 (+0.31) | 4.47 (+0.34) | 4.42 (+0.34) | 4.27 (+0.36) |
Key Findings¶
- The Decisive Value of Explicit Legal Structure: Moving from free-form anticipation (A-GuideRAG) to the constrained IFH in SCALE yields substantial gains (>0.3 points) in Precedent Relevance, Precedent Application, and Judicial Alignment, showing that unconstrained query expansion frequently drifts into peripheral legal topics.
- Structural Scaffolding Compensates for Model Scale: Smaller 8B models equipped with SCALE rival or outperform unassisted 70B parameter models (e.g., Qwen3 8B with SCALE scores 4.56 on Legal Issue ID and 4.36 on Precedent Application, surpassing Llama-3.1 70B on all baselines). Explicit domain structure bridges the capacity gap of smaller weights.
- Automated Evaluators Fail on Nuanced Precedent Application: Comparing GPT-4.1 (LLM-as-a-Judge) with expert scores reveals weak-to-moderate Spearman correlations (\(\rho = 0.11\) for Precedent Relevance, \(\rho = 0.09\) for Precedent Application). While automated judges evaluate surface fluency and structural coherence reasonably well (\(\rho \approx 0.31-0.34\)), they struggle to verify whether a legal precedent genuinely governs a specific factual dispute.
Highlights & Insights¶
- From Free-Form Generation to Graph Traversal: Instead of expecting an LLM to hallucinate or intuit the governing legal standard, SCALE frames legal anticipation as an issue-factor traversal problem over an explicit doctrinal topology.
- Tapping Authoritative Secondary Sources: Utilizing ECHR Case Law Guides as structured intermediate knowledge solves the high noise and heterogeneity problems endemic to searching raw judicial opinion databases.
- Actionable Framework for Specialized High-Stakes Domains: Demonstrates that for domains requiring precise doctrinal rigor, externalizing conceptual frameworks into structured hierarchies is far more reliable and cost-effective than scaling parameter sizes alone.
Limitations & Future Work¶
- Dependency on Curated Case Law Guides: SCALE leverages human-curated ECHR guides; in jurisdictions without such guides, automated hierarchy extraction from judicial corpora (e.g., via LexGenie) may introduce noise that propagates downstream.
- High Cost of Expert Legal Evaluation: Evaluating 888 multi-page legal reasoning traces requires verified domain expertise, placing fundamental limits on evaluation scalability.
- Expansion to Long-Form Legal QA: The authors highlight that issue-aligned, factor-specific retrieval can be naturally generalized to statutory long-form question answering and cross-jurisdictional legal analysis.
Related Work & Insights¶
- vs LeCoPCR & LexKeyPlan: These methods introduce concept-guided and keyphrase planning for legal retrieval, but lack global hierarchical constraints; SCALE tethers retrieval directly to article-specific Issue–Factor Hierarchies.
- vs CourtReasoner & LegalChainReasoner: While earlier frameworks simulate multi-agent courtroom debate or decompose statutes into premise-conclusion triplets, SCALE directly addresses the challenge of identifying latent factual factors and linking them to controlling precedent.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Elegant formalization of case law guides into an Issue–Factor Hierarchy for anticipatory retrieval.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous expert evaluation of 888 reasoning traces alongside an informative LLM-as-a-Judge correlation study.
- Writing Quality: ⭐⭐⭐⭐⭐ Highly articulate, technically rigorous, and deeply grounded in both legal doctrine and NLP methodology.
- Value: ⭐⭐⭐⭐⭐ Provides an exemplary blueprint for high-stakes, knowledge-intensive legal AI systems.