1. Bibliographic Information

1.1. Title

The title of the paper is "LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI". The central topic is the development of a framework to audit and mitigate hallucinations in artificial intelligence systems specifically deployed for legal contract analysis. The paper focuses on moving beyond aggregate error rates to understand the specific types and directions of errors (omission vs. invention) to ensure trustworthy deployment.

1.2. Authors

The authors of the paper are Lalit Yadav and Akshaj Gurugubelli. While the provided text does not explicitly list their affiliations, the nature of the work suggests a background in Natural Language Processing (NLP), Legal AI, and Machine Learning research.

1.3. Journal/Conference

The paper was published on arXiv (Cornell University's open-access archive for preprints in physics, mathematics, computer science, etc.). The identifier is arXiv:2606.18021. As a preprint published in June 2026, it represents the latest cutting-edge research before potential peer-reviewed publication in a top-tier AI or Law conference (such as NeurIPS, ACL, or ICAIL).

1.4. Publication Year

The paper was published in 2026.

1.5. Abstract

The paper addresses the issue of high hallucination rates (approx. 52%) in AI systems used for legal workflows. It argues that aggregate metrics conceal critical details about where errors concentrate and their direction (omission vs. invention). The authors present LegalHalluLens, an auditing framework with three main components:

1. Typed Hallucination Profiles: Categorizing hallucinations into four legal claim types: numeric, temporal, obligation/entitlement, and factual.

2. Risk Direction Index (RDI): A scalar metric that captures the bias of a model towards omitting information versus inventing it.

3. Typed Debate Pipeline: A multi-agent mitigation system calibrated using the specific failure profiles identified by the audit.

   Key findings include a significant gap (38-40 percentage points) in error rates between obligation/numeric claims and temporal claims, which aggregate reporting hides. The framework demonstrates that systems with identical aggregate rates can have opposite risk biases. The debate pipeline successfully reduces fabricated detections by 45%, enabling a smaller open model (4B parameters) to match commercial API performance.

1.6. Original Source Link

The official source link is: https://arxiv.org/abs/2606.18021 The PDF link is: https://arxiv.org/pdf/2606.18021v1 The publication status is a preprint.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to solve is the lack of actionable, granular error signals in current Legal AI evaluations. While it is known that AI systems hallucinate (generate false information) at high rates (~52%), standard reporting relies on aggregate metrics. These averages mask the specific nature of the errors, leaving compliance officers unable to assess the true risk of deployment.

In legal contexts, not all errors are equal. Inventing a liability cap (invention) creates a false sense of security, while missing a non-compete qualifier (omission) might render a clause unenforceable. Prior research has established that hallucinations are not uniform but has not sufficiently addressed the specific context of contract extraction or provided a single, deployment-comparable metric to capture the direction of errors (omission vs. invention). The paper's entry point is the hypothesis that "typed" profiling (classifying claims by verification difficulty) and "direction-aware" auditing can provide the necessary signal for trustworthy AI.

2.2. Main Contributions / Findings

The paper makes three primary contributions:

1. Typed Hallucination Profiles: The authors introduce a consistent failure ordering across four claim categories: ${numeric, obligation} >> factual >= temporal$. They reveal a massive within-model gap of approximately 38–41 percentage points between high-risk claims (numeric/obligation) and low-risk claims (temporal), which is completely invisible when looking at aggregate scores.
2. Risk Direction Index (RDI): A novel signed scalar metric that distinguishes between models that tend to "invent" conditions (positive RDI) versus those that tend to "omit" them (negative RDI). This metric separates systems that appear identical under aggregate evaluation.
3. Calibrated Multi-Agent Debate: A mitigation pipeline where the behavior of agents (specifically the "Skeptic" and safety gates) is calibrated based on the specific failure profile of the model (diagnosed in step 1) rather than using generic prompts. This approach reduces fabricated detections by 45% and allows a small open model to outperform commercial APIs on composite scores.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand this paper, one must be familiar with several key concepts:

• Large Language Models (LLMs): Advanced AI models (like GPT, Gemini, Llama) trained on vast amounts of text to generate human-like responses. They are the backbone of modern Legal AI.
• Hallucination: In the context of LLMs, hallucination refers to the phenomenon where the model generates content that is nonsensical, unfaithful to the source material, or factually incorrect.
• Contract Extraction: A specific Natural Language Processing (NLP) task where the goal is to identify and extract specific clauses or data points (e.g., "Liability Cap," "Termination Date") from a legal contract document.
• CUAD Dataset: The Contract Understanding Atticus Dataset is a benchmark dataset used in the paper. It consists of 510 commercial contracts annotated by legal experts with 41 different clause types. It serves as the "ground truth" or "oracle" for evaluation.
• Multi-Agent Debate: A technique where multiple AI instances (agents) interact, critique, and debate each other's outputs to arrive at a more accurate or reliable final answer. This paper uses a specific variation of a "Graph of Thoughts" or "Debate" framework.

3.2. Previous Works

The paper positions itself against several clusters of prior work:

• Legal Hallucination Typology: Research by Dahl et al. (2024), Hou et al. (2024), and Magesh et al. (2025) established that legal hallucinations vary in type and severity. However, the authors note that these works did not focus specifically on the contract extraction setting or collapse directional character into a single scalar metric.
• General Hallucination Benchmarks: Benchmarks like FActScore (Min et al., 2023), HaluBench, and HalluLens (Bang et al., 2025) measure factual precision but generally lack the stratification by specific legal claim types (numeric vs. obligation) required for legal risk assessment.
• Debate-Based Mitigation: Previous studies (e.g., Du et al., 2024, Liang et al.,) have explored multi-agent debate to improve factuality. The key differentiator in this paper is that prior work tuned debate prompts generically across all error types, whereas this paper calibrates the debate agents specifically to the diagnosed failure modes of the model.

3.3. Technological Evolution

The field has evolved from general-purpose LLM evaluation to domain-specific benchmarks (like LegalBench and CUAD). Initially, evaluations focused on simple accuracy. The focus then shifted to detecting hallucinations. This paper represents the next step: diagnostic auditing. Instead of just asking "How often does it fail?", it asks "How does it fail, in which direction, and on what specific types of claims?" It moves from static evaluation to calibrated mitigation, where the evaluation results directly dictate the architecture of the fix (the debate pipeline).

3.4. Differentiation Analysis

The core difference between this paper's approach and prior methods is the calibration loop.

• Prior Work: Generic evaluation $\rightarrow$ Generic Mitigation. (e.g., "Use a Skeptic agent to check everything.")
• This Paper: Typed Evaluation (identifying specific weak points) $\rightarrow$ Calibrated Mitigation (e.g., "Use a Skeptic agent that specifically asks about numeric values because the audit showed the model fails there.") Additionally, the introduction of the Risk Direction Index (RDI) provides a unique scalar for comparing models based on their "bias" (invention vs. omission), a dimension often ignored in standard accuracy/F1-score comparisons.

4. Methodology

4.1. Principles

The methodology is built on the principle of Audit-Diagnosis-Calibration.

1. Audit: Rigorously evaluate a model on a legal corpus to generate a "Typed Hallucination Profile" and "Risk Direction Index."
2. Diagnosis: Analyze the profile to identify specific weaknesses (e.g., "This model invents numbers" or "This model drops obligations").
3. Calibration: Configure a multi-agent debate pipeline where the "Skeptic" agent is programmed to challenge the model specifically on its diagnosed weaknesses, and the safety gates are tuned to the specific error direction (omission vs. invention).

4.2. Core Methodology In-depth (Layer by Layer)

Step 1: Defining the Typed Hallucination Profile

The first step involves partitioning the legal claims into four distinct categories based on their primary verification challenge. This categorization is crucial because different types of claims require different verification logic and carry different risks.

• Numeric: Claims involving values (e.g., liability caps). Verification requires checking for exactness and units.

• Temporal: Claims involving dates and durations. Verification requires verbatim string matching.

• Obligation/Entitlement: Claims about duties or rights (e.g., "Supplier shall..."). Verification is complex, requiring checking modal verbs (shall/may), conditions, and carve-outs.

• Factual: Identity claims (e.g., governing law). Verification requires resisting parametric priors (what the model "knows" generally) in favor of the specific document text.

  For a given model $M$ evaluated on a corpus $\mathcal{D}$, the model outputs a clause extraction or a "not present" decision for each claim type $c_i$. The performance is measured using $\mathrm{Hal_{TP}}$ (Hallucination rate among True Positives), which measures content correctness conditional on detection.

The paper defines the Within-Model Typed Gap to quantify the disparity in performance across these categories. The formula is presented as follows:

$$\mathrm{Gap}(M)=\max_{c_{i}}\mathrm{Hal_{TP}}(M,c_{i})-\min_{c_{i}}\mathrm{Hal_{TP}}(M,c_{i})$$

Where:

• $\mathrm{Gap}(M)$ is the difference in hallucination rates between the worst-performing and best-performing claim categories for model $M$.

• $\max_{c_{i}}$ and $\min_{c_{i}}$ find the maximum and minimum values across the set of claim categories $c_i \in \{\mathrm{numeric}, \mathrm{temporal}, \mathrm{obligation}, \mathrm{factual}\}$.

• $\mathrm{Hal_{TP}}(M, c_i)$ is the hallucination rate for model $M$ on category $c_i$.

  A large $\mathrm{Gap}(M)$ indicates that the model's reliability varies wildly depending on what you ask it to extract, rendering an aggregate score misleading.

Step 2: Calculating the Risk Direction Index (RDI)

The second component of the methodology addresses the direction of the error. A model can fail by "inventing" a clause that doesn't exist (False Positive/Invention) or "missing" a clause that does exist (False Negative/Omission). Furthermore, within a detected clause, it can add extra conditions or miss required ones.

The authors use an external judge to label errors. Two specific labels are critical for RDI:

• missing_condition: The model omitted a qualifier present in the ground truth.

• extra_condition: The model asserted a qualifier absent from the source.

  The Risk Direction Index (RDI) is calculated using the following formula:

$$\mathrm{RDI}(M)=\frac{\mathrm{p_{extra}}(M)-\mathrm{p_{missing}}(M)}{100}$$

Where:

• $\mathrm{RDI}(M)$ is the Risk Direction Index for model $M$.
• $\mathrm{p_{extra}}(M)$ is the percentage of contradicted findings labeled as extra_condition.
• $\mathrm{p_{missing}}(M)$ is the percentage of contradicted findings labeled as missing_condition.

Interpretation:

• Positive RDI: The model tends to invent or overstate (add conditions that aren't there).

• Negative RDI: The model tends to omit or understate (drop conditions that are there).

• Near Zero: The model is balanced in its error direction.

  This scalar allows procurement officers to compare models based on risk profile. For example, in a compliance workflow where missing an obligation is catastrophic, a model with a Positive RDI (inventing things) might be safer than one with a Negative RDI (hiding things), because invented details are easier for a human reviewer to spot than missing ones.

Step 3: The Typed Debate Pipeline

The third component is the mitigation pipeline. This is a multi-agent system designed to correct the baseline extraction. The pipeline is a state machine with six roles, operating in phases.

The following figure (Figure 3 from the original paper) illustrates the architecture of this pipeline:

Phase 1: Debate

1. Skeptic: This agent issues typed challenge questions. Crucially, these questions are not generic; they are derived from the Typed Hallucination Profile from Step 1.
   • Example: If the profile shows high failure on numeric claims, the Skeptic asks: "Is this exact value verbatim in the source, or is it a common prior assumption?"
   • Example: If the profile shows high failure on obligation claims, the Skeptic asks: "Are all carve-outs and exceptions captured?"
   • Example: If the profile shows high failure on factual claims, the Skeptic asks: "Is this information from the document or external knowledge?"
2. Supporter: Defends the extraction using only verbatim quotes from the contract.
3. Routing & Re-extractor: If the Skeptic identifies a structural error (wrong clause extracted), a Re-extractor node fires to re-run the extraction from the source. This targets structural errors distinct from content errors.
4. Arbiter: If agents disagree after rounds, the Arbiter resolves the deadlock using a conservative policy (preserving the baseline unless contrary evidence is strong).

Phase 2: Independent Verification 5. Verifier: This agent searches the contract independently to check if the extracted definition fits the context.

Phase 3: Judging with Safety Gates 6. Judge: Reads the transcript, Verifier report, and Arbiter assessment to make final decisions. This step uses Asymmetric Structural Gates calibrated to the RDI. * Addition Gate (Absent $\rightarrow$ Present): Requires both Verifier confirmation and debate consensus. This blocks fabricated additions (high False Positive risk). * Deletion Gate (Present $\rightarrow$ Absent): Blocked if the Verifier confirms presence. This prevents the pipeline from being too conservative and erasing correct findings (high False Negative risk).

    The asymmetry encodes the risk profile: if the model is prone to invention (Positive RDI), the Addition Gate is strict. If the model is prone to omission (Negative RDI), the Deletion Gate is restrictive.

5. Experimental Setup

5.1. Datasets

The experiments utilize the CUAD v1.0 dataset.

• Source: CUAD (Contract Understanding Atticus Dataset) by Hendrycks et al. (2021).
• Scale: 510 commercial contracts.
• Annotations: 41 expert-annotated clause types per contract.
• Characteristics: It provides a complete ground-truth oracle. Every model output can be verified against the contract text without external knowledge.
• Usage:
  • Experiment 1: Evaluated on all 510 contracts.
  • Experiment 2: Evaluated on a 120-contract matched subset for the mitigation study.
• Why Chosen: The oracle-bounded nature is critical for calculating hallucination rates accurately. The authors mapped the 41 clause types to their 4 categories (Numeric, Temporal, Obligation, Factual) to enable the typed profiling.

5.2. Evaluation Metrics

The paper employs a comprehensive set of metrics to evaluate detection accuracy and content quality.

1. False Acceptance Rate (FAR)

• Conceptual Definition: This metric quantifies the rate at which the model "hallucinates" the presence of a clause that is actually absent in the source document. It measures the propensity for invention.
• Mathematical Formula: $$\mathbf{FAR}=\mathrm{FP}/(\mathrm{FP}+\mathrm{TN})$$
• Symbol Explanation:
  • $\mathrm{FP}$: False Positives (Instances where the model said a clause is present, but it was not).
  • $\mathrm{TN}$: True Negatives (Instances where the model correctly said a clause is absent).

2. False Rejection Rate (FRR)

• Conceptual Definition: This metric quantifies the rate at which the model misses clauses that are actually present in the source document. It measures the propensity for omission.
• Mathematical Formula: $$\mathbf{FRR}=\mathrm{FN}/(\mathrm{FN}+\mathrm{TP})$$
• Symbol Explanation:
  • $\mathrm{FN}$: False Negatives (Instances where the model missed a clause that was present).
  • $\mathrm{TP}$: True Positives (Instances where the model correctly detected a clause).

3. Accuracy (Acc)

• Conceptual Definition: The overall proportion of correct predictions (both presence and absence) out of all instances.
• Mathematical Formula: $$\mathbf{Acc}=(\mathrm{TP}+\mathrm{TN})/N$$
• Symbol Explanation:
  • $N$: Total number of instances.

4. Hallucination Rate on True Positives ($\mathrm{Hal_{TP}}$)

• Conceptual Definition: This measures content quality conditional on detection. Of the clauses the model found, what fraction had incorrect content (e.g., wrong number, missing condition)? This isolates content errors from detection errors.
• Mathematical Formula: $$\mathbf{Hal_{TP}}=\mathrm{contradicted}/\mathrm{TP}$$
• Symbol Explanation:
  • $\mathrm{contradicted}$: The number of True Positives where the extracted content contradicted the ground truth.

5. General Hallucination Rate ($\mathrm{Hal_{Gen}}$)

• Conceptual Definition: A stricter metric that counts everything the model emitted as a clause. It penalizes both content contradictions (wrong content) and fabrications (hallucinated presence). This is the primary metric for evaluating the mitigation pipeline.
• Mathematical Formula: $$\mathbf{Hal_{Gen}}=(\mathrm{contradicted}+\mathrm{FP})/(\mathrm{TP}+\mathrm{FP})$$
• Symbol Explanation:
  • The numerator includes all content errors plus all false positive detections.
  • The denominator is the total number of clauses the model generated.

6. Joint Equality (JEq)

• Conceptual Definition: An end-to-end correctness metric measuring the fraction of instances where the model's output exactly matches the ground truth (correct detection AND correct content).
• Mathematical Formula: $$\mathbf{JEq}=\mathrm{supported}/(\mathrm{TP}+\mathrm{FN})$$
• Symbol Explanation:
  • $\mathrm{supported}$: The number of instances where the model's output was fully supported by the ground truth.

5.3. Baselines

The paper compares the proposed method against several strong baselines:

• Commercial APIs: gemini-3-flash and gpt-5.2. These represent state-of-the-art proprietary models.
• Open Models: qwen3-32b (32.8B parameters) and llama-3.3-70b (70B parameters). These represent large open-source models.
• Mitigation Baseline: For the debate pipeline experiment, the baseline is gemma-4-26B-A4B (a Mixture-of-Experts model with 4B active parameters). This model was specifically chosen because it had the worst baseline score on the subset, ensuring that any improvement is due to the intervention.

6. Results & Analysis

6.1. Core Results Analysis

The experiments reveal three critical insights:

1. Aggregate Metrics are Misleading: All four models tested in Experiment 1 fell within a narrow 6 percentage point band for $\mathrm{Hal_{TP}}$ (50.9% - 56.5%). A compliance officer looking at these numbers would see no meaningful difference between the systems.
2. Typed Gap is Massive: When stratifying by claim type, a consistent pattern emerged: ${numeric, obligation} >> factual >= temporal$. The hallucination rate for numeric/obligation claims was 64.8–74.3%, while for temporal claims it was only 29.0–35.1%. This represents a ~40 percentage point gap hidden by the average.
3. Direction Matters: Two models, qwen3-32b and gpt-5.2, had nearly identical aggregate scores (~52%). However, qwen3-32b had a negative RDI (omission-heavy), while gpt-5.2 had a positive RDI (invention-heavy). This distinction is vital for legal risk assessment.
4. Calibrated Debate Works: The typed debate pipeline significantly improved the gemma-4-26B-A4B model. It reduced fabricated detections by 45% and moved the model from last place to first place on the composite score leaderboard, matching commercial APIs.

6.2. Data Presentation (Tables)

Experiment 1: Aggregate Metrics

The following are the results from Table 1 of the original paper, showing that aggregate metrics provide little separation between models.

Experiment 1: Typed Hallucination Profiles

The following are the results from Table 2 of the original paper, revealing the significant gap between claim types.

The following figure (Figure 1 from the original paper) visualizes this data, highlighting the grey band of aggregate rates versus the high failure rates of numeric/obligation claims.

该图像是一个条形图，展示了510份合同中不同类型的幻觉率。数值型和义务型的幻觉率达到64.8%至74.3%，而时间型的幻觉率为29.0%至35.1%。整体的幻觉率范围为50.9%至56.5%。

Experiment 1: Risk Direction Index (RDI)

The following are the results from Table 3 of the original paper, showing the directional bias of each model.

The following figure (Figure 2 from the original paper) illustrates the error direction, showing that qwen3-32b predominantly misses conditions (red bars) while gpt-5.2 invents them (green bars).

该图像是一个图表，展示了在不同模型中缺失条件和额外条件的错误方向。不同模型的缺失条件出现在红色条形中，额外条件的发明则在绿色条形中。qwen3-32b模型在缺失条件方面的错误为21.1%，而gpt-5.2模型的额外条件发明为21.0%，两者的合规风险方向指标（RDI）分别为-0.202和0.161。

Experiment 2: Matched-Subset Comparison

The following are the results from Table 4 of the original paper, comparing the calibrated debate pipeline against baselines.

The "gemma-debate" row shows the performance of the proposed method, achieving the best composite score (2.4) compared to the commercial API "gpt-5.2" (2.6) and the base model "gemma-base" (5.2).

6.3. Ablation Studies / Parameter Analysis

Per-Category Gains

The authors analyzed the per-type deltas resulting from the debate pipeline. The following figure (Figure 4 from the original paper) shows these gains:

该图像是图表（图4），展示了不同类型的检测增益情况。它显示了义务类型的增益最大（$ext{ΔFAR} = -8.2, ext{ΔHal}_{ ext{Gen}} = -6.3$），其后是事实类型（$ext{ΔFAR} = -5.8$），而时间类型的增益变化基本保持不变（$+0.6$ pp），与基线预测一致。

The analysis confirms the calibration hypothesis:

• Obligation & Factual: Showed the largest gains ($\Delta \mathrm{FAR} = -8.2$ and -5.8 respectively). This matches the diagnosis that these were high-failure categories where the Skeptic's specific challenges (checking for carve-outs and external knowledge) were most effective.
• Temporal: Showed essentially no change ($+0.6$ pp). This is consistent with the baseline finding that temporal claims were already the lowest-hallucination type and hard to fabricate verbatim.

RDI Shift Analysis

The authors also measured how the debate pipeline affected the error direction. The following figure (Figure 5 from the original paper) shows the RDI shift for the gemma-4-26B-A4B model:

该图像是条形图，展示了应用了有针对性的辩论管道后，gemma-4-26B-A4B模型的风险方向指数（RDI）变化。不同类别的RDI显示出显著修正，尤其是义务类别最大的变化为 (-0.078 ightarrow -0.014)，显示出接近平衡的状态。

• Observation: The obligation category RDI shifted significantly from -0.078 (omission-heavy) to -0.014 (near-balanced).
• Interpretation: The typed Skeptic challenges (specifically targeting missing conditions and carve-outs) successfully corrected the model's bias towards dropping critical obligation details. This validates the "Calibrated" aspect of the methodology: the intervention fixed the specific directional bias identified in the audit.

Robustness Analyses

The paper includes several robustness checks in Appendix E:

• Per-Run Variance: Standard deviations across three runs were small ($\le 2.4$ pp), confirming the typed ordering is stable.
• Composite Rank Sensitivity: The gemma-debate model ranked first under 4 out of 5 different weighting schemes (FP-heavy, Recall-heavy, Halluc-only, Detection-only), proving the result is robust to metric prioritization.
• Debate Pipeline Overhead: The pipeline was efficient, with a mean of 1.12 rounds (max 2). 87.9% of cases finished in Round 1. This indicates the benefit comes from the quality of the calibrated challenge, not just extended computation time.

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper successfully demonstrates that aggregate hallucination metrics are insufficient for legal AI deployment. By introducing Typed Hallucination Profiles and the Risk Direction Index (RDI), the authors uncover a ~40 percentage point gap in reliability between different legal claim types and separate models with identical aggregate scores but opposite failure biases. Furthermore, they prove that these diagnostics can be used to calibrate a multi-agent debate pipeline, significantly reducing fabrications (by 45%) and enabling a small open model to outperform commercial APIs. The framework provides a actionable path for direction-aware procurement and accountable agent design.

7.2. Limitations & Future Work

The authors explicitly acknowledge several limitations:

1. Scope of Evidence: The numerical results apply specifically to CUAD-style English-US commercial contracts. Whether the failure ordering transfers to other jurisdictions or document types (e.g., patents, court filings) remains an empirical question.
2. Judge Dependence: All metrics flow through a single external LLM judge (gemini-2.5-flash). While the judge was held fixed and independent, the lack of human-validated labels for the entire dataset means small RDI differences should be interpreted with caution.
3. Context Window: Experiments assumed full-document context. For contracts exceeding context windows, retrieval-augmented generation (RAG) would introduce additional failure modes not measured here.
4. Mitigation Scope: The debate pipeline was shown to reduce fabrications (False Positives) effectively but was less effective at correcting content errors within present clauses (content contradictions), consistent with prior findings that LLMs struggle to self-correct reasoning.

7.3. Personal Insights & Critique

This paper offers a highly valuable and pragmatic contribution to the field of Trustworthy AI. It moves beyond the "leaderboard mentality" of aggregate scores and addresses the nuanced reality of legal risk.

• Innovation: The operationalization of "Error Direction" into a single scalar (RDI) is particularly elegant. It transforms a qualitative observation ("this model makes stuff up") into a quantitative, comparable metric.

• Calibration Loop: The idea of using the diagnosis to configure the cure (the debate pipeline) is a powerful methodological advancement. It moves away from "one-size-fits-all" mitigation strategies.

• Potential Issues: The reliance on an LLM-as-a-Judge is a double-edged sword. While scalable, it introduces a dependency on another model's biases. Future work should definitely focus on validating these RDI scores against human legal experts.

• Transferability: The "Typed Profile" concept is highly transferable. In medical AI, for instance, one could distinguish between "Dosage Hallucinations" vs "Symptom Hallucinations," as the risks of inventing a dosage are likely higher than omitting a symptom description.

• Improvement: The paper notes the pipeline struggles with content correction. A future iteration could integrate a "Critic" agent that specifically compares the extracted text span against the source character-by-character for high-stakes numeric fields, rather than relying on semantic debate alone.

  Overall, LegalHalluLens provides a rigorous framework for the necessary next step in AI adoption: moving from "Does it work?" to "How does it fail, and is that failure acceptable for my specific risk profile?"