Dynamic Adjudication Template (DAT)

• Dynamic Adjudication Template is a framework that manages decision-making by constraining LLM outputs to a predefined, auditable candidate pool.
• It employs techniques like retrieval-constrained value-set completion, graph-based execution, and iterative draft–verify–revise loops to enhance precision and transparency.
• The approach integrates multiple adjudication methods, enabling legal and multimodal applications to balance recall and rule-governed selection in high-stakes domains.

Searching arXiv for papers relevant to “Dynamic Adjudication Template” and adjacent adjudication frameworks. arxiv_search(query="\"Dynamic Adjudication Template\" OR adjudication LLM constrained retrieval legal adjudication", max_results=10) “Dynamic Adjudication Template” is best treated as an Editor’s term for a family of adjudication architectures in which an LLM or composite system is restricted to bounded selection, verification, or revision over explicit alternatives, rather than unconstrained generation. In the current literature, that pattern is instantiated most clearly by retrieval-constrained value-set completion, multi-agent legal judgment workflows, deterministic graph-based legal execution, and progressive multimodal dispute resolution (Mukherjee, 22 Jun 2026); (Liao et al., 13 Mar 2026); (Sójka et al., 4 May 2026); (Wu et al., 18 Mar 2026). The term is also a source of acronymic ambiguity: some arXiv papers using “DAT” concern data-architecture tooling, hybrid retrieval weighting, or productivity measurement rather than adjudication (Abughazala et al., 30 Jan 2025); (Abughazala et al., 2023); (Hsu et al., 29 Mar 2025); (Beller et al., 14 Mar 2025).

1. Terminological scope and acronymic ambiguity

The phrase “Dynamic Adjudication Template (DAT)” is not a universal term of art across the cited literature. In particular, the DATCloud paper states that it does not define or use the term “Dynamic Adjudication Template (DAT),” and the 2023 smart-cities paper uses DAT in the sense of a “Data Architecture Modeling Tool,” not adjudication (Abughazala et al., 30 Jan 2025); (Abughazala et al., 2023). A separate 2025 retrieval paper uses DAT for “Dynamic Alpha Tuning,” and a 2025 software-engineering paper uses DAT for “Diff Authoring Time” (Hsu et al., 29 Mar 2025); (Beller et al., 14 Mar 2025).

Use of “DAT”	Meaning in source	Relation to adjudication
DAT / DATCloud	Data Architecture Modeling Tool / DATCloud (Abughazala et al., 30 Jan 2025); (Abughazala et al., 2023)	Not an adjudication construct
DAT	Dynamic Alpha Tuning (Hsu et al., 29 Mar 2025)	Adjudicates between retrievers
DAT	Diff Authoring Time (Beller et al., 14 Mar 2025)	Productivity metric, not adjudication

This ambiguity matters because the adjudicative sense of DAT is synthesized from several papers rather than introduced as a single canonical framework. The closest direct formulation appears in clinical value-set authoring, where a constrained LLM adjudicator operates over an auditable candidate pool (Mukherjee, 22 Jun 2026). Closely related forms appear in legal judgment prediction, neuro-symbolic legal execution, and ride-hailing responsibility disputes, where the common structure is bounded decision-making over explicit evidence, rules, or candidate sets (Liao et al., 13 Mar 2026); (Sójka et al., 4 May 2026); (Wu et al., 18 Mar 2026).

2. Formal foundations of adjudication

The formal substrate predates LLMs. “Diversity and Adjudication” models the outputs of multiple implementations as a non-empty bag,

$BAG == \{ b : VALUE \rightarrow \mathbb{N} \mid \ran b \neq \{0\} \},$

and defines an adjudication operator as

$ADJOP == BAG \leftrightarrow VALUE.$

Within that framework, adjudication is the problem of obtaining a single result from multiple, potentially different outcomes (Boiten, 2015).

The paper then enumerates several adjudicators with distinct semantic assumptions. Majority voting is a partial function defined when some value occurs strictly more than half the time:

$$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$

Other adjudicators include First-Past-the-Post, Greatest Lower Bound, PLUBF for flat domains with a distinguished failure value, the median, and weak choice, where any observed value may be returned (Boiten, 2015).

These definitions establish the durable core of the DAT idea: adjudication is not generation but a rule-governed mapping from a set, multiset, or structured collection of alternatives to an accepted outcome. This suggests that modern DATs inherit classical concerns—unanimity, majority, permutation-invariance, partiality, and fault masking—even when the alternatives are no longer N-version program outputs but retrieved codes, legal drafts, graph clauses, or queried evidence.

3. Candidate-pool adjudication and tool arbitration

The clearest contemporary DAT instance is RASC+, which studies “a stage-wise alternative in which candidate-pool construction is optimized for recall and a constrained LLM adjudicator is optimized for candidate selection” (Mukherjee, 22 Jun 2026). For a target value set $v$, Stage 1 constructs a candidate pool $C_v$; Stage 2 selects a subset $\hat{Y}_v \subseteq C_v$. Candidate-pool recall is defined as

$$\mathrm{PoolRecall}(v) = \frac{|Y_v \cap C_v|}{|Y_v|}.$$

On the full 3,744-value-set RASC test split, retrieval raises candidate-pool recall from 0.553 to 0.730, and replacing the stage-2 selector with blinded GPT-5 adjudication raises full-test macro F1 from 0.287 to 0.549 and held-out-publisher macro F1 from 0.233 to 0.533 (Mukherjee, 22 Jun 2026).

The decisive property is constraint. The LLM is never asked to generate new codes; it can only choose among supplied candidates. Operationally, the prompt enumerates candidates with candidate_ids, requires strict JSON, and any output not matching candidate IDs can be discarded. Because some pools exceed context limits, pools are deterministically chunked into subsets of $\le 1{,}200$ candidates, adjudicated independently, and unioned afterward (Mukherjee, 22 Jun 2026). The system therefore preserves an auditable output space while still using an LLM for semantic discrimination.

A narrower but related form of adjudication appears in “Dynamic Alpha Tuning for Hybrid Retrieval in Retrieval-Augmented Generation,” where an LLM scores dense and BM25 top-1 results and computes a per-query mixture coefficient $\alpha(q)$ (Hsu et al., 29 Mar 2025). The fusion rule is

$$R(q, d) = \alpha(q) \cdot \tilde{S}_{\text{dense}}(q, d) + \bigl(1 - \alpha(q)\bigr) \cdot \tilde{S}_{\text{BM25}}(q, d),$$

with $ADJOP == BAG \leftrightarrow VALUE.$0 derived from LLM effectiveness scores for the two retrievers. With GPT-4o as judge, DAT reaches $ADJOP == BAG \leftrightarrow VALUE.$1 and $ADJOP == BAG \leftrightarrow VALUE.$2 on SQuAD, and $ADJOP == BAG \leftrightarrow VALUE.$3 and $ADJOP == BAG \leftrightarrow VALUE.$4 on DRCD, outperforming fixed-weight hybrid retrieval (Hsu et al., 29 Mar 2025). This suggests a broader DAT interpretation: adjudication may select among candidates inside a pool, or among tools that construct or rank that pool.

4. Collegial and directive-informed legal adjudication

VERDICT instantiates DAT as a self-refining collaborative multi-agent legal process. Its inference framework is formalized as

$ADJOP == BAG \leftrightarrow VALUE.$5

where $ADJOP == BAG \leftrightarrow VALUE.$6 is the input space, $ADJOP == BAG \leftrightarrow VALUE.$7 the set of agents, $ADJOP == BAG \leftrightarrow VALUE.$8 the intermediate states, and $ADJOP == BAG \leftrightarrow VALUE.$9 the controller over agents and states (Liao et al., 13 Mar 2026). The agent roles are specialized and complementary: Court Clerk Agent for fact structuring, Judicial Assistant Agent for legal retrieval, Case-handling Judge Agent for opinion drafting, Adjudication Supervisor Agent for supervisory verification, and Presiding Judge Agent for final adjudication (Liao et al., 13 Mar 2026).

The dynamic core is the draft–verify–revise loop. At iteration $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$0, the Judge produces a draft, the Supervisor returns $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$1 together with natural-language feedback $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$2, and the feedback is appended to the history before the next draft (Liao et al., 13 Mar 2026). The process terminates on Pass or when the maximum turn count is reached. This converts adjudication from a one-shot classifier into a traceable revision process.

VERDICT’s second dynamic element is Hybrid Jurisprudential Memory (HJM), grounded in the Micro-Directive Paradigm. A Micro-Directive is represented as

$$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$3

where $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$4 is directive content, $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$5 is a confidence or lifespan score, $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$6 are supporting and opposing precedents, and $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$7 links the directive to statute identifiers (Liao et al., 13 Mar 2026). The framework is evaluated on CAIL2018 and CJO2025, the latter using a strict future time-split composed only of judgments after January 1, 2025 (Liao et al., 13 Mar 2026). This suggests a DAT as a reusable, structured circuit in which roles, intermediate representations, verification signals, and evolving directive memory are all first-class components.

5. Neuro-symbolic and multimodal evidentiary templates

A different legal DAT form appears in “Accurate Legal Reasoning at Scale,” where a high-capacity model is used once to translate legal text into Deterministic Autonomous Contract Language (DACL), a typed graph intermediate representation, and runtime adjudication is delegated to deterministic graph execution (Sójka et al., 4 May 2026). Given a query $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$8 and facts $$\forall x : VALUE; b : BAG \cdot \mathbf{MV}(b) = x \;\Leftrightarrow\; 2 * b(x) > size(b).$$9, the runtime pipeline routes to relevant clause identifiers $v$0, executes

$v$1

and then synthesizes a natural-language explanation from the value $v$2 and trace $v$3 (Sójka et al., 4 May 2026). On 400 test events across four real contracts, the DACL Agent reaches 99.5% overall accuracy and 98% on Logistics-MSA; it uses $v$4M tokens versus $v$5M for GPT-5.2 Medium, and the system reduces compute costs by over 90% in high-volume workflows (Sójka et al., 4 May 2026). Here, the template is dynamic chiefly through semantic routing, clause versioning, and auditable execution traces rather than free-form reasoning at every step.

RideJudge supplies a multimodal, quasi-judicial DAT for ride-hailing responsibility disputes. Its pipeline combines SynTraj for visual grounding, Adaptive Context Optimization for rule and precedent condensation, a Chain-of-Adjudication for progressive evidentiary inquiry, and Ordinal-Sensitive Reinforcement Learning for graded liability calibration (Wu et al., 18 Mar 2026). SynTraj generates 12,585 image–caption pairs linking abstract liability concepts to concrete trajectory patterns, and CoA yields 14,582 high-confidence adjudication chains after filtering out cases that disagree with human labels or are marked “Ambiguous” (Wu et al., 18 Mar 2026).

The Chain-of-Adjudication is explicitly interactive. An Adjudicator reasons over textual order information and candidate rules, requests needed visual facts through <map>...</map>, a Visual Analyst answers with <answer>...</answer>, and a Reasoning Refiner restructures the dialogue into four stages—Information Analysis, Visual Evidence Integration, Rule Grounding, and Comprehensive Adjudication—before emitting <reason>, <judge>, and <result> (Wu et al., 18 Mar 2026). RideJudge-8B achieves 88.41% accuracy overall and 91.86% on the Appeal benchmark (Wu et al., 18 Mar 2026). Its ordinal label space is written as

$v$6

and the answer reward gives full reward for exact matches and distance-sensitive reward for near-miss predictions (Wu et al., 18 Mar 2026). This is a DAT in which the action space is bounded not only by rule subsets and evidence queries, but also by an ordered severity ontology.

6. Limits, misconceptions, and open problems

A recurrent misconception is that any paper using “DAT” concerns adjudication. That is false for DATCloud and the smart-cities DAT work, both of which concern model-driven data architecture and explicitly do not define a “Dynamic Adjudication Template” artifact (Abughazala et al., 30 Jan 2025); (Abughazala et al., 2023). DATCloud uses “pre-defined templates and reusable components” to speed up architecture modeling, but it contains “no explicit notion of ‘adjudication’” and no runtime decision mechanism under that name (Abughazala et al., 30 Jan 2025). The adjudicative DAT is therefore not an acronymic constant but a cross-paper pattern.

Across actual adjudication systems, the limitations are substantial. RASC+ is benchmark-based, chunking prevents cross-chunk comparison, and token-level cost and latency were not measured (Mukherjee, 22 Jun 2026). VERDICT improves transparency and temporal generalization, but its multi-agent loops increase inference time, its evaluation is in a Civil Law setting, and directive quality depends on “consistent, labeled, and validated cases” (Liao et al., 13 Mar 2026). DACL’s expressivity is limited to arithmetic, first-order logic, range-based lookups, and temporal validity; ambiguous, open-textured provisions remain outside the executable graph, and human legal-engineer review is mandatory before production use (Sójka et al., 4 May 2026). RideJudge addresses interpretability and evidentiary rigor, but privacy-sensitive data, historical bias, fairness, robustness, and governance remain live issues in quasi-judicial deployment (Wu et al., 18 Mar 2026).

Taken together, these papers suggest four recurrent properties of a DAT. First, the output space is bounded: by candidate IDs, clause graphs, rule subsets, or explicitly queried evidence. Second, intermediate objects are auditable: provenance features, Pass/Reject feedback, DACL traces, or structured reason sections. Third, adjudication is iterative rather than one-shot: retrieve then select, draft then verify then revise, or ask for evidence before ruling. Fourth, adaptation is externalized into explicit mechanisms such as memory, retrieval, versioned clauses, or rule calibration rather than hidden inside unconstrained generation. In that sense, the modern DAT is less a single model class than a disciplined adjudication pattern for high-stakes, evidence-constrained decision systems.