Judicial Continuation Discourse

• Judicial Continuation Discourse is defined as the evolving, iterative process by which legal reasoning is carried forward, refined, and tested across cases and time.
• It integrates methodologies from sentence-level rhetorical role annotation to multi-agent systems, ensuring coherence between facts, inferences, and judicial outcomes.
• Empirical benchmarks show that leveraging continuity in legal discourse can enhance precedent retrieval, doctrinal extension, and accuracy in judgment prediction.

Judicial Continuation Discourse denotes a conception of adjudication in which legal reasoning is carried forward, tested, and reformulated across time rather than emitted as a static one-shot decision. In the most explicit formulation, it is “the ongoing, iterative, panel-like formation, verification, and refinement of legal rationales across cases and time,” but related work extends the notion to the continuity of a judgment’s internal rhetoric, the transmission of argument from oral hearing to written opinion, and the case-by-case extension of doctrine under institutional and temporal constraints (Liao et al., 13 Mar 2026, Held et al., 2023). Taken together, this suggests an umbrella concept for how courts and court-like AI systems preserve coherence while moving from facts to reasons to outcomes under precedent, hierarchy, and evolving standards (Caputo, 8 May 2026).

1. Conceptual Scope and Core Meaning

In legal judgment prediction, Judicial Continuation Discourse is presented as a direct alternative to static classification. VERDICT defines Legal Judgment Prediction as a legally grounded mapping from tokenized facts $s_d=\{w_1^d,\dots,w_{l_d}^d\}$ to a triplet $j=(a,c,t)$ of applicable law articles, charges, and penalty term,

$\mathcal{F}: s_d \to (a,c,t),$

and then reframes that mapping as a collegial and temporally evolving reasoning process rather than a purely discriminative output layer (Liao et al., 13 Mar 2026).

At the level of written judgments, the concept acquires a document-internal meaning. HKJudge defines judicial continuation discourse as the way a judgment carries forward “what the court finds, how it reasons, and what it ultimately rules” across sentences and sections. Its two-tier schema makes this progression explicit through 26 sentence-level rhetorical roles grouped under Fact, Inference, Result, and Other, plus span-level annotations for charge, imprisonment term, and fine (Xuan et al., 4 Jun 2026). In this formulation, continuity is not only temporal; it is also rhetorical and structural.

At the procedural level, LaCour! shifts the locus of continuation to the passage from oral hearing into later written law. The corpus makes it possible to study how judges’ questions, parties’ arguments, and rhetorical structures in European Court of Human Rights hearings persist, are reframed, or are contested in majority judgments and separate opinions (Held et al., 2023). Continuation here includes both uptake and transformation: dialogic, multilingual oral exchanges become formal, structured written reasons.

A further institutional meaning appears in the Critical Discourse Analysis of Data Intellectual Property Rights in China. There, “judicial continuation discourse” is the author’s label for the way local courts extend established intellectual property norms to data disputes in the absence of clear legislation, building doctrine case by case and circulating local reasoning upward across judicial hierarchies (Lou, 23 Sep 2025). This use emphasizes doctrinal extension rather than merely argumentative persistence.

2. Jurisprudential Foundations

A major theoretical line links Judicial Continuation Discourse to jurisprudence and AI alignment. “Alignment as Jurisprudence” argues that judges and AI systems share a common structure because both must predict and shape future decisions under open-textured language. In that account, Dworkin’s “fit” and “justification” correspond to principled continuation of law, while Sunstein’s analogical reasoning and “incompletely theorized agreements” describe continuation through precedents and low-level principles. The same essay maps jurisprudential tools—rules, principles, precedent—onto alignment tools—policies, constitutions, case-based grounding—and treats next-token generation as a formally similar act of continuation that can be normatively guided (Caputo, 8 May 2026).

Common-law analysis makes the carriers of continuation empirically measurable. In the ECtHR study of precedent, the core question is whether continuity is transmitted primarily by prior arguments or by prior facts. Using conditional mutual information,

$\mathrm{MI}(O;H\mid F)=H(O\mid F)-H(O\mid H,F),$

the paper reports that precedent arguments share 0.38 nats of information with current case outcomes, whereas precedent facts share 0.18 nats. This supports Halsbury’s view overall, while also showing that fact-based continuity can dominate in specific doctrinal areas where legal concepts are less straightforward (Valvoda et al., 2021).

A complementary foundation comes from Bayesian-narrative theories of fact-finding. The reconciliation of Legal Probabilism with narration-based approaches models adjudication as a dynamic interplay among competing narratives that are evaluated for coherence, completeness, explanatory power, plausibility, specificity, and resilience. Conviction under Classical Legal Probabilism is stated as

$\text{Convict if } P(G\mid E)\ge T,$

but the paper adds narrative-sensitive conditions such as domination, resiliency, and reasonable doubt, so that adjudicative continuation is not exhausted by a single posterior number (Urbaniak, 2017). This suggests that continuation discourse is simultaneously probabilistic and narrative: the record evolves, and so do the stories that organize it.

3. Formal Representations of Continuity

Recent work formalizes Judicial Continuation Discourse at multiple granularities. At the court-system level, “When Precedents Clash” enriches classifier-based legal case-based reasoning with court hierarchy, binding relations, and temporal ordering. An organization is represented as $\mathrm{Org}=(\mathrm{Courts},H,B)$, with $H$ a strict partial order and $B$ the binding relation, while the temporal relation $T$ is a total preorder on cases. Conflict resolution is then defined through hierarchical and temporal selectors such as

$$\mathrm{Best}_{TH}(\tilde S)=\mathrm{Best}_T(\mathrm{Best}_H(\tilde S)),$$

allowing overruling and per incuriam to be formalized as exceptions to precedential constraint (Florio et al., 2024).

At the sentence-pair level, discourse relation modeling treats continuation as a local semantic relation between adjacent units. The U.S. court-transcript study defines five top-level relations—Elaboration, Redundancy, Citation, Shift in View, and No Relation—and places continuation-like phenomena primarily under Elaboration, which includes several CST relations such as Follow-up, Description, Historical Background, and Attribution (Ratnayaka et al., 2018). This is a narrow but operational notion of continuation: one sentence carries forward the same argumentative thread as its predecessor.

At the sentence-sequence level, HKJudge turns continuity into an annotation schema for entire judgments. Its 26 rhetorical roles distinguish, among other functions, F0-charge, F2-event, I1-case_law, I2-ordinance, I6-assertion, R, and R-other; 1.97% of sentences are multi-labeled, largely because appellate judgments embed lower-court discourse within current reasoning (Xuan et al., 4 Jun 2026). The schema is designed precisely to expose transitions such as Fact $j=(a,c,t)$0 Inference $j=(a,c,t)$1 Result.

At the document-tree level, the copyright-damages study models opinions using Rhetorical Structure Theory. Judicial continuity is represented through nuclei, satellites, and multinuclear relations such as Sequence, List, and Joint, so that an opinion can be parsed into a hierarchical tree $j=(a,c,t)$2 whose local relations nest into larger macro-structures connecting infringement analysis, liability analysis, and damages analysis (Lo et al., 14 Jan 2026). The paper’s Tree-of-Discourse linearization makes a specific span’s contribution to the overall reasoning path explicit.

A related but distinct representational layer appears in CuRIAM, which focuses on metalanguage rather than rhetoric alone. Its annotation schema covers Focal Term, Definition, Metalinguistic Cue, Direct Quote, Indirect Quote, Legal Source, Language Source, Named Interpretive Rule, Example Use, and Appeal to Meaning (2305.14719). In this framework, continuation is enacted through repeated revoicing of legal texts, interpretive canons, and authoritative definitions.

4. Workflow Realizations in Legal AI

Judicial Continuation Discourse has been operationalized most explicitly in multi-agent legal AI. AgentsCourt mirrors the court process through three sequential modules: Court Debate Simulation, Legal Information Retrieval, and Judgment Refinement. It treats “Determine facts” as an immutable shared state, preserves transcripts as persistent argumentative memory, and uses a judge assistant to perform case-type prediction, BM25 rough retrieval over the appropriate sub-corpus, dense re-ranking with BGE-Large embeddings, and extraction of legal grounds from the top-5 precedents (He et al., 2024). In this architecture, continuity is procedural: debate, retrieval, and judgment are connected rather than isolated.

VERDICT generalizes the same intuition into a self-refining collegial panel. Its workflow is defined on a DAG,

$j=(a,c,t)$3

with Court Clerk, Judicial Assistant, Case-handling Judge, Adjudication Supervisor, and Presiding Judge agents participating in a draft–verify–revise loop with explicit Pass/Reject control signals. Its Hybrid Jurisprudential Memory stores Standard Precedents and evolving Micro-Directives, and retrieves memory units with a score combining legal applicability, graph topology, and semantic similarity:

$j=(a,c,t)$4

Validated reasoning traces are archived and distilled into Micro-Directives, linking today’s verification trajectory to tomorrow’s guidance (Liao et al., 13 Mar 2026).

The public-assault example in VERDICT illustrates this design. Facts include “public BBQ stall,” “venting anger,” and “Minor Injury.” The Assistant retrieves a directive, “Public Place + Venting Anger → Art. 293 (Picking Quarrels).” The Case Judge initially drafts Art. 234 (Intentional Injury). The Supervisor rejects the draft on the ground that public location and provocation indicate social-order violation outweighing personal injury, and the system revises to Art. 293 (Liao et al., 13 Mar 2026). Continuation is therefore not only memory accumulation; it is correction through controlled iteration.

A more general systems perspective appears in the requirements analysis of generative AI for judicial reasoning. That paper argues that the decisive stages for any legal reasoner are Rule and Application under IRAC, and maps RAG, multi-agent systems, neuro-symbolic AI, structured reasoning, and output verification to those tasks. It stresses that cross-jurisdictional framework selection, ratio decidendi extraction, open-textured standards, conflict resolution, and burden-of-proof application are the hardest stages because they require justified legal choices rather than fluent completion (Linna et al., 26 Aug 2025).

5. Corpora, Benchmarks, and Empirical Evidence

A substantial empirical literature now studies Judicial Continuation Discourse through dedicated corpora and workflow benchmarks.

Resource	Scope	Main use
HKJudge	4,000 Hong Kong criminal judgments; $j=(a,c,t)$5292,240 sentence–tag pairs; $j=(a,c,t)$66.5M tokens	Rhetorical role classification; legal element extraction
LaCour!	154 ECHR hearings; 2,161,833 tokens; over 267 hours	Hearing-to-judgment continuity; judge-question analysis
SimuCourt / JudicialKB	420 Chinese judgment documents; first and second instance; large legal KB	Trial and appellate continuity in agent simulation
TriBench-Ko	1,414 binary items across 31 task-risk pairs	Risk-aware evaluation in judicial workflows
CAIL2018 / CJO2025	134,739 and 8,199 LJP samples	Temporal generalization and legal judgment prediction

HKJudge provides the first sentence-level expert-annotated Hong Kong judgment corpus, with ten legal linguistics annotators and reported inter-annotator agreement of $j=(a,c,t)$7. It benchmarks encoders, open-source LLMs, and commercial LLMs on rhetorical role classification and legal element extraction. For rhetorical role classification, Qwen-2.5-72B with fine-tuning reaches Accuracy $j=(a,c,t)$8 and Macro-F1 $j=(a,c,t)$9, while commercial systems perform better, with Claude-Opus-4 at Accuracy $\mathcal{F}: s_d \to (a,c,t),$0 and Macro-F1 $\mathcal{F}: s_d \to (a,c,t),$1, and Gemini-2.5-Pro at Accuracy $\mathcal{F}: s_d \to (a,c,t),$2 and Macro-F1 $\mathcal{F}: s_d \to (a,c,t),$3. For legal element extraction, Qwen-2.5-72B with fine-tuning reaches Accuracy $\mathcal{F}: s_d \to (a,c,t),$4 and Macro-F1 $\mathcal{F}: s_d \to (a,c,t),$5, while Claude-Opus-4 and Gemini-2.5-Pro are around Accuracy $\mathcal{F}: s_d \to (a,c,t),$6 and $\mathcal{F}: s_d \to (a,c,t),$7, Macro-F1 $\mathcal{F}: s_d \to (a,c,t),$8 and $\mathcal{F}: s_d \to (a,c,t),$9 respectively (Xuan et al., 4 Jun 2026).

LaCour! makes oral-to-written continuity directly measurable. It links 154 ECHR hearings to corresponding HUDOC judgments and opinions, includes sentence-level timestamps, role and language labels, and enables judge-level alignment between oral questioning and subsequent opinion writing. In a preliminary experiment, Fisher’s exact test rejects independence between asking a question and having an authored opinion with $\mathrm{MI}(O;H\mid F)=H(O\mid F)-H(O\mid H,F),$0, and the odds of a judge having an opinion are 1.55 times higher if they asked a question. For 193 English question–opinion pairs from the same judge, Sentence-BERT ranking yields precision@1 = 0.16, indicating detectable but modest semantic alignment (Held et al., 2023).

AgentsCourt reports gains from explicit procedural continuity. On legal grounds, it achieves improvements of 8.6% F1 in the first-instance setting and 9.1% F1 in the second-instance setting over best baselines, with ablations showing degradation from 0.203 in the full system to 0.171 without court simulation and 0.145 without the knowledge base. In the retrieval pipeline, cause-of-action consistency between the current case and the retrieved top-1 precedent improves from 62% after rough retrieval to 85% after dense re-ranking (He et al., 2024).

VERDICT evaluates continuation under temporal shift. CAIL2018 contains 134,739 samples with 99 law articles, 115 charges, and 11 prison-term categories, while CJO2025 contains 8,199 samples, 67 law articles, 73 charges, and 11 prison-term categories, with a strict future split after January 1, 2025 and postdating the knowledge cutoffs of DeepSeek-V3 and Qwen2.5. On CAIL2018, VERDICT reports Law Article Acc 85.35, Charge Acc 82.40, and Penalty Term Acc 39.76; on CJO2025, it reports Law Article Acc 90.56, Charge Acc 85.84, and Penalty Term Acc 45.68. The ablation removing $\mathrm{MI}(O;H\mid F)=H(O\mid F)-H(O\mid H,F),$1 lowers CJO2025 Law Article accuracy from 90.56 to 88.46, indicating that Micro-Directives contribute to temporal generalization (Liao et al., 13 Mar 2026).

TriBench-Ko evaluates continuity under deployment risks rather than pure task accuracy. It covers jurisprudence summarization, precedent retrieval, legal issue extraction, and evidence analysis, and assesses hallucination, omission, statutory misapplication, demographic bias, overcompliance, prompt sensitivity, nondeterminism, and adjudicative overreach. The hardest region is Precedent Retrieval × Omission, with mean F1 $\mathrm{MI}(O;H\mid F)=H(O\mid F)-H(O\mid H,F),$2. Overall macro F1 is highest for gpt-5.4 at 0.835, followed by gpt-5.4-mini at 0.781 and Qwen3.5-9B at 0.771 (Lee et al., 5 May 2026).

6. Limits, Risks, and Future Directions

The literature converges on a common limitation: continuity is easier to state than to guarantee. Judicial Requirements for Generative AI argues that legal reasoning becomes most difficult at Rule selection and Application because these stages require justified doctrinal choices, analogical evaluation, conflict resolution, and correct burden-of-proof handling. The paper further states that prompts and model selection become discoverable methodological choices, so transparency must extend to the generative process rather than only the final answer. Its proposed near-term role for AI is therefore dual: a high-volume assistant for simple, repetitive cases and a “sparring partner” for human experts in complex matters (Linna et al., 26 Aug 2025).

Risk-sensitive benchmarks sharpen that caution. TriBench-Ko shows that omission, especially in precedent retrieval and summarization, is often more dangerous than overt hallucination because outputs can be directionally correct yet materially incomplete. It also documents demographic bias, overcompliance, prompt sensitivity, nondeterminism, and adjudicative overreach, the last of which occurs when a model crosses from assistance into court-like normative resolution (Lee et al., 5 May 2026).

System-specific limitations remain substantial. VERDICT notes that the expert Case Judge is a 7B model due to resource constraints, that multi-agent iteration increases inference latency, that the framework is evaluated in a civil-law setting, and that confidence decay in directive evolution is heuristic. AgentsCourt identifies bias in precedent selection, hallucination of statutes, retrieval noise, long-debate drift from “Determine facts,” and possible underuse of first-instance records in second-instance reasoning if prompts are not crafted carefully (Liao et al., 13 Mar 2026, He et al., 2024).

A broader normative controversy concerns how rules and cases should be combined. “Alignment as Jurisprudence” warns against over-reliance on highly general principles, which can leave interpretation to pretraining priors, and against over-reliance on curated case repositories, which can narrow pluralism or miss salient features. Its proposed guardrails are jurisprudentially explicit: elevate meta-principles protecting rights, require public reason-giving at the principle and case levels, expose rationales to audit, and retain human final authority in courts (Caputo, 8 May 2026).

This suggests that Judicial Continuation Discourse is becoming a design criterion rather than merely a descriptive label. In current research, a system qualifies as continuity-aware not because it produces long explanations, but because it preserves traceable links among facts, authorities, reasoning steps, revisions, and institutional boundaries across time. The central open problem is therefore not the generation of more legal text; it is the construction of reasoning processes that remain coherent under doctrinal change, procedural review, and adversarial scrutiny.