PaperlayerView exact revision on arXiv

Source-linked AI summary

LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI

Lalit Yadav, Akshaj Gurugubelli

arXiv:2606.18021v1cs.AIcs.CLcs.LGcs.MA

TL;DR

Aggregate legal-AI hallucination rates do not show which claim types fail or whether errors reflect omission or invention. LegalHalluLens audits typed failures and calibrates debate accordingly, measuring a ∼40 pp typed gap and reducing fabricated detections by 45%.

  • Problem

    Existing legal-AI evaluation lacks evidence on systematic hallucination differences across claim types and error directions, despite consequential model-selection decisions.

  • Method

    LegalHalluLens combines typed hallucination profiles, the Risk Direction Index, and a debate pipeline with agents targeting measured failure modes.

  • Results

    A consistent ∼40 pp gap separates obligation/numeric from temporal claims, while typed debate reduces fabricated detections by 45%.

  • Takeaways & Limitations

    Typed profiles and RDI distinguish failure patterns that aggregate evaluation misses, supporting direction-aware legal-AI deployment decisions.

  • Takeaways & Limitations

    Results are limited to 510 English-US commercial contracts from CUAD, and generalisation across jurisdictions and document types remains unverified.

Abstract

from arXiv · show

AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving compliance officers without an actionable signal for trustworthy deployment. We present LegalHalluLens, an auditing framework with three components: typed hallucination profiles across four legally-motivated claim categories (numeric, temporal, obligation/entitlement, factual) over CUAD (Hendrycks et al., 2021); a Risk Direction Index (RDI) that reduces omission-versus-invention bias to a single deployment-comparable scalar; and a typed debate pipeline calibrated to both magnitudes and directions. Across 510 contracts and 249,252 clause-level instances we measure a within-model gap of approximately 38-40 pp between obligation/numeric and temporal claims that aggregate reporting hides, and show that two systems with matched 52% rates can carry opposite RDIs. The debate pipeline reduces fabricated detections by 45% with per-category gains tracking the diagnosis, matching commercial APIs with a substantially smaller backbone (4B active parameters). Typed profiles and RDI surface failure modes that aggregate metrics hide; we further show these diagnostics serve as calibration inputs for multi-agent debate pipelines, where Skeptic challenges and asymmetric gates targeted at measured failure modes outperform generically-tuned debate. The framework supports direction-aware procurement, accountability, and agent design for legal AI deployed in the wild.

1. Introduction

LegalHalluLens argues that trustworthy legal AI requires typed, direction-aware hallucination auditing rather than aggregate error rates alone. It introduces typed profiles, the Risk Direction Index, and calibrated multi-agent debate to diagnose and mitigate failures in contract extraction.

  • Motivation: Legal AI is deployed in consequential workflows where model outputs and model selection create real legal exposure.Applications include contract review, compliance monitoring, regulatory reporting, and due diligence.
  • Problem: A 52% aggregate hallucination rate conceals which legal clauses fail and whether errors invent risks or omit obligations.The passages illustrate invented liability caps and silently dropped non-compete qualifiers as asymmetric legal harms.
  • Typed hallucination profiles: Approximately 38–41 pp separates numeric and obligation hallucinations from factual and temporal claims across four architecturally diverse models.The ordering is {numeric, obligation} ≫ factual ≥ temporal and is not observable under aggregate reporting.
  • Risk Direction Index: The Risk Direction Index encodes omission-versus-invention bias as a signed, deployment-actionable scalar across typed claim categories.It is designed to differentiate directional error character that aggregate rates cannot capture.
  • Calibrated multi-agent debate: 45% fewer fabricated detections were achieved by a six-role calibrated debate pipeline, enabling a 4B-active open model to match commercial APIs on composite score.Skeptic challenges and Add/Delete gate asymmetries were derived from diagnosed failure magnitudes and directions; the model ranked first under 4 of 5 weighting schemes.

2. Related Work

Prior work establishes that legal hallucinations vary by mode and persist despite retrieval augmentation, while general benchmarks and debate methods often lack claim-type stratification. LegalHalluLens builds on this literature by arguing that high-stakes agent design should be conditioned on typed failure modes and directional risk.

  • Legal hallucinations and benchmarks: Legal hallucination research identifies varied error modes, reports model-dependent rates of 58%–88%, and finds that retrieval augmentation does not eliminate hallucinations.Dahl et al. develop a legal hallucination typology; Hou et al. propose fine-grained gap categories; Magesh et al. examine RAG in commercial legal AI tools.
  • General hallucination benchmarks and debate-based mitigation: FActScore, HaluBench, HalluLens, and PHANTOM measure factual precision without stratifying claims by type, while multi-agent debate has been studied for factuality.The passage also situates debate in theoretical work on inference-time scaling.
  • Agent design in high-stakes deployment: Recent debate systems tune Skeptic prompts, gate thresholds, and aggregation rules generically, whereas LegalHalluLens conditions challenges on observed failure modes and gate asymmetry on directional risk.The framework positions typed profiles and RDI as inputs for deployment-oriented agent design.

3. Background

Legal claims differ in how they must be verified, motivating four source-checkable categories: numeric, temporal, obligation/entitlement, and factual. The framework evaluates extraction decisions with confusion-matrix outcomes and distinguishes hallucination metrics by whether they condition on detections or cover all emitted clauses.

  • Verification regimes: Numeric and temporal claims can be checked by direct comparison with source values or verbatim strings, whereas obligation claims require more complex verification.The passage illustrates numeric caps and termination dates as direct-comparison cases and indemnification language as an obligation example.
  • Claim taxonomy: Legal claims are grouped into numeric, temporal, obligation/entitlement, and factual categories according to their primary verification challenge.The taxonomy is intended to transfer across legal extraction tasks whenever model claims can be checked against a source.
  • Evaluation setup: For each document–claim-type pair, the model outputs either a clause extraction or a “not present” decision evaluated against CUAD using TP, FP, FN, and TN outcomes.TP denotes correctly detected present clauses; FP fabricated absent clauses; FN missed present clauses; TN correctly absent clauses.
  • Hallucination metrics: HalTP measures contradicted content among detected clauses, isolating extraction-content errors conditional on finding the clause.It is the primary signal for typed profiles in Experiment 1.
  • Hallucination metrics: HalGen measures wrong content across every emitted clause, counting both contradictions and fabrications rather than conditioning only on detections.This makes HalGen stricter in scope than HalTP.

4. Method

The method combines typed hallucination profiling, the Risk Direction Index (RDI), and a debate pipeline whose mitigation decisions are informed by the evaluation outputs. It stratifies clause-level errors by claim type, captures omission-versus-invention direction, and applies targeted agent challenges and asymmetric repair gates.

  • Method components: Three components comprise the method: typed hallucination profiles and RDI for evaluation, plus a typed debate pipeline for mitigation.The debate pipeline is informed by the outputs of the two evaluation procedures.
  • Typed hallucination profile: HalTP(M, ci) is reported by numeric, temporal, obligation, and factual claim categories to expose within-model variation hidden by aggregate averages.The typed gap identifies models whose hallucination rates vary substantially across categories with different deployment consequences.
  • Risk Direction Index: RDI uses missing-condition and extra-condition labels to distinguish omission-heavy from invention-heavy contradicted findings.Positive RDI values indicate invention-heavy failure, while negative values indicate omission-heavy failure.
  • Typed debate pipeline: The debate state machine uses six agent roles, including a Skeptic targeting failure modes, a quote-constrained Supporter, a Re-extractor, and an Arbiter.The Re-extractor reruns extraction from the source when the wrong clause was selected, rather than debating an unrepairable answer.
  • Typed debate pipeline: 62–71% of content contradictions are scope errors, motivating separate structural re-extraction and asymmetric gates for additions and deletions.The Addition Gate requires Verifier confirmation and debate consensus, while the Deletion Gate blocks removal when the Verifier confirms presence.

5. Experiments

The experiments audit legal clause extraction on CUAD using four verification-oriented claim categories, then test typed debate mitigation with held-out models and a matched subset. A single external judge evaluates both experiments against CUAD ground truth using strict precision, modality, polarity, and exception-preservation criteria.

  • Dataset and categories: CUAD v1.0 supplies 510 commercial contracts, 41 expert-annotated clause types, and complete contract-only ground truth for verification.The 41 clause types are mapped to Numeric, Temporal, Obligation/Entitlement, and Factual categories by primary verification challenge.
  • Dataset and categories: The central typed-gap comparison uses Obligation/Entitlement (n=27) versus Temporal (n=6), while Numeric and Factual results provide supporting evidence because their category sizes are small.The categories target modal drift and condition loss, implied deadline inference, threshold fabrication and unit mismatch, and outside-knowledge injection, respectively.
  • Experimental design: Experiment 1 benchmarks gemini-3-flash, gpt-5.2, qwen3-32b, and llama-3.3-70b at temperature = 0 with identical structured-JSON prompts.The commercial APIs are gemini-3-flash and gpt-5.2; qwen3-32b and llama-3.3-70b are open models with 32.8B and 70B parameters.
  • Experimental design: Experiment 2 tests typed debate mitigation with gemma-4-26B-A4B, a held-out Mixture-of-Experts backbone with 4B active parameters, selected as the worst matched-subset baseline.The design keeps mitigation separate from benchmarking so improvements are attributable to the intervention rather than a stronger starting point.
  • Evaluation: A single external gemini-2.5-flash judge at temperature = 0 scores both experiments against CUAD ground truth using five criteria and returns supported/contradicted verdicts with mismatch labels.The criteria are numeric precision, temporal precision, modality match, polarity match, and exception/carve-out preservation; this judge produces every reported HalTP, HalGen, and RDI value.
  • Experimental design: 249,252 clause-level instances result from three independent Experiment 1 runs across all 510 contracts, while Experiment 2 uses a 120-contract matched subset with 4,920 nominal opportunities.Actual Experiment 1 exports are 62,580, 62,689, 61,536, and 62,447 for gemini-3-flash, gpt-5.2, qwen3-32b, and llama-3.3-70b, respectively.

6. Results: Typed Hallucination Profiles (Experiment 1)

Aggregate HalTP compresses materially different failure profiles: models cluster within 50.9–56.5%, while numeric and obligation hallucinations reach 64.8–74.3% versus 29.0–35.1% for temporal claims. Directional diagnostics further distinguish systems with nearly identical aggregate rates, showing that deployment choice depends on claim type and the legal task’s error asymmetry.

  • Aggregate metrics: Four architecturally distinct models fall within a 6 pp aggregate HalTP band of 50.9–56.5%, offering compliance officers no actionable deployment signal.The comparison includes two commercial APIs, a 32B open model, and a 70B open model.
  • Typed profiles: Numeric and obligation claims hallucinate at 64.8–74.3% across models, whereas temporal claims hallucinate at 29.0–35.1%, producing an approximately 38–41 pp within-model gap.The ordering {numeric, obligation} ≫ factual ≥ temporal holds for every model without exception.
  • Typed profiles: No model dominates every category: qwen3-32b leads numeric at 66.8% and temporal at 29.0%, gpt-5.2 leads obligation at 64.8%, and gemini-3-flash leads factual at 36.0%.Gemini-3-flash also leads end-to-end JEq at 46.9%.
  • Deployment implications: A low aggregate error profile can conceal weak consequential extraction: llama-3.3-70b has FAR 7.7% and Acc 89.0%, but numeric FRR 52.8% and numeric JEq 12.1%.This shows that conservative abstention is not necessarily a safe fallback for deployments centered on numeric clauses.
  • Directional profiles: qwen3-32b and gpt-5.2 report essentially identical HalTP rates of 52.1% and 51.8% but fail in opposite directions: omission versus invention.Scope errors dominate at 62–71%; RDI captures only the clearly directional portion yet cleanly separates the systems, and the appropriate profile depends on the legal task’s error asymmetry.

7. Results: Calibrated Mitigation (Experiment 2)

Experiment 2 shows that calibrating typed debate to Experiment 1’s failure profile reduces fabrication most in high-failure categories. The intervention improves composite ranking and corrects obligation-direction bias, while leaving content contradictions largely unchanged.

  • Calibration: Skeptic challenges target per-type failure profiles, while asymmetric gates encode the measured FAR > FRR risk asymmetry.The pipeline uses claim-type-specific challenges, contract-quote defenses, independent verification, and conservative adjudication.
  • Mitigation effects: False-positive extractions drop from 524 to 287 (−45%), while content contradictions move only 642 to 641 (−0.2%).This indicates fabrication filtering rather than reliable correction of what present clauses say.
  • Per-type results: Obligation ∆FAR = −8.2, factual ∆FAR = −5.8, numeric ∆FAR = −3.6, and temporal HalGen is essentially unchanged (+0.6 pp).The observed ordering matches the pre-specified prediction that gains would concentrate in obligation and factual claims and be smallest for temporal claims.
  • Comparative performance: Gemma-debate improves from Score 5.2 to 2.4 and clears gpt-5.2 at 2.6 on composite score.The comparison holds the underlying model fixed, isolating the intervention’s effect.
  • Direction correction: Obligation RDI shifts from −0.078 to −0.014, moving gemma-4-26B-A4B from omission-heavy to near-balanced.Challenges target missing conditions, dropped carve-outs, and scope loss, addressing the omission bias identified in Experiment 1.

8. Discussion

The discussion argues that typed gaps and RDI make aggregate legal-AI evaluation more actionable for governance, while emphasizing that reported profiles apply to CUAD-style English US commercial contracts. The auditing method may transfer to other verifiable legal tasks, but practitioners should commission task-specific audits rather than reuse CUAD-derived thresholds.

  • Deployment and governance implications: ∼40 pp typed gap separates temporal-claim failure rates of 29–35% from 65–74% for liability-threshold and obligation-scope claims.Aggregate HalTP reporting conceals this within-model disparity.
  • Deployment and governance implications: Two systems with identical HalTP scores can have opposite risk profiles, which RDI expresses as a single comparable number.This distinction supports direction-aware compliance evaluation.
  • Scope: The reported typed profiles and RDI values apply to CUAD-style English US commercial contracts, and transfer to other document types remains empirically unresolved.The paper does not establish that the observed failure ordering transfers beyond this scope.
  • Scope: The auditing method can instantiate typed profiles and RDI for any legal task with a verifiable oracle, but the resulting numbers will differ.Practitioners should commission task-specific audits instead of applying CUAD-derived thresholds to new contexts.

9. Conclusion

LegalHalluLens shows that aggregate hallucination metrics conceal substantial category gaps and opposite risk profiles, while typed debate reduces fabricated detections in line with prior diagnosis.

  • 9. Conclusion: ~40 pp (38.0–40.6 pp across models) separates obligation/numeric from temporal hallucinations across 249,252 clause-level instances, a gap aggregate evaluation conceals.The measurement spans four models and four claim categories.
  • 9. Conclusion: Two models with matched HalTP carry opposite risk profiles, operationalised by the Risk Direction Index.The Risk Direction Index distinguishes directionality that matched aggregate rates do not reveal.
  • 9. Conclusion: 45% reduction in fabricated detections comes from a typed debate pipeline whose per-category gains track the prior diagnosis.The result supports using typed diagnostic profiles to target debate interventions.

Limitations

The study’s evidence is bounded by its CUAD contract sample, full-document context assumptions, and limited experimental replication. Its HalTP, HalGen, and RDI results also depend on a single LLM judge, so small RDI differences require caution.

  • Dataset and scope: Results cover 510 English-US commercial contracts from CUAD, and generalisation across jurisdictions and document types remains unverified.The typed failure ordering is consistent across four architectures, but this does not establish broader generalisation.
  • Context assumptions: All experiments assume full-document context, while retrieval-augmented variants for longer contracts may introduce additional, unmeasured failure modes.These retrieval-related failure modes are described as orthogonal to those measured in the study.
  • Experimental replication: Experiment 2 uses one run with one backbone, gemma-4-26B-A4B, on a 120-contract subset.This limits the replication and backbone diversity supporting that experiment’s findings.
  • Evaluation dependence: All HalTP, HalGen, and RDI numbers flow through a single LLM evaluation judge, gemini-2.5-flash.The judge is fixed and independent of the extractors, but the absence of human-validated labels means small RDI differences should not be over-interpreted; examples include gpt-5.2 (+0.161) versus qwen3-32b (−0.202).

Impact Statement … E. Robustness Analyses

The paper frames LegalHalluLens as a diagnostic for direction-aware deployment and calibrated agent design, while limiting claims to CUAD-style English US commercial contracts. Its appendices specify strict judging, exhaustive extraction, typed skeptic challenges, category mappings, and robustness analyses using the experimental data.

  • Impact Statement: Typed profiles should inform, not replace, qualified human review because even the best configuration contradicted detected clause contents 58.6% of the time.The framework is presented as supporting model comparison, risk-aware deployment, and mitigation design rather than providing clearance for autonomous use.
  • Impact Statement: The RDI makes deployment trade-offs legible: compliance workflows may prefer positive or near-zero RDI, while legal operations may prefer the opposite profile.The distinction reflects different costs for missed obligations and false positives.
  • Impact Statement: The authors recommend per-deployment typed-profile remeasurement and human review of all flagged obligation and numeric clauses before autonomous deployment.They warn that calibrated pipelines could create an appearance of compliance review while masking the 40+ pp typed gap.
  • Impact Statement: The reported numerical results apply to CUAD-style English US commercial contracts, while failure rates should be re-measured for each new deployment context.The methodology is described as extensible to legal tasks with a verifiable source.
  • A. Judge Prompt: The external evaluation judge used one identical prompt across all four extraction backbones in Experiment 1 and all six configurations in Experiment 2.The judge determines semantic equivalence between an AI answer and ground truth for the same clause and contract, using strict precision criteria.
  • B. Extraction Prompt (abbreviated): The extraction prompt restricts models to contract text, requires outputs for all 41 CUAD clause types, and asks for operative spans including conditions, exceptions, notice periods, and cross-references.A self-check targets additional conditions, numeric thresholds, and cross-references; outputs are JSON arrays and temperature is 0.
  • B.1 Numeric Clause Definitions (verbatim): Numeric clause definitions distinguish explicit caps, awarded liquidated damages, minimum commitments, usage thresholds, and pricing discretion from related but nonqualifying concepts.The definitions explicitly exclude consequential-damages exclusions from Cap On Liability, denial of liquidated damages from Liquidated Damages, minimum quotas from Volume Restriction, and payment limits from Price Restrictions.
  • C. Typed Skeptic Challenge Questions: Typed skeptic questions are derived from each type’s dominant Experiment 1 failure mode, targeting numeric precision, obligation modality and conditions, temporal exactness, and factual grounding.Numeric checks cover values, units, caps, floors, and qualifiers; obligation checks cover modal verbs, triggers, exceptions, and scope; temporal checks distinguish explicit dates and units; factual checks require contract-stated entities.

E.1 Per-Run Variance (Experiment 1) … E.6 Debate Pipeline Overhead

The appendix validates that typed hallucination patterns and RDI directionality are stable, while examining ranking robustness, missing-row causes, obligation heterogeneity, and debate overhead. These analyses show persistent category separation, contract-correlated processing failures, and targeted debate changes without extensive deliberation.

  • E.1 Per-Run Variance (Experiment 1): Within-model typed gaps remain stable across runs, with standard deviations small relative to the 38.0–40.6 pp gap.All within-model typed gaps remain ≥36 pp at the 1-SD bound.
  • E.2 RDI Bootstrap CIs by Category: The dominant Obligation category preserves deployment-relevant directional separation in pooled 95% bootstrap confidence intervals.The intervals use 2,000 resamples over all runs pooled.
  • E.2 RDI Bootstrap CIs by Category: Obligation RDI values range from +0.220 for gpt-5.2 to −0.198 for llama-3.3-70b, with gemini-3-flash at +0.018 and qwen3-32b at −0.181.The reported 95% CIs are [+0.207, +0.234], [+0.004, +0.031], [−0.210, −0.186], and [−0.194, −0.168], respectively.
  • E.3 Composite Rank Sensitivity: gemma-debate ranks first under 4 of 5 weighting schemes, while gpt-5.2 leads under recall-heavy weighting.The intervention’s improvement over gemma-base, ranked 5–6 in every scheme, is robust to weighting.
  • E.4 Missing-Row Attribution: For qwen3-32b, missing rows are contract-correlated: 5 contracts are incomplete in all 3 runs and 58 in any run, yielding a persistent fraction of 8.6%.The failures occur under temperature = 0 API calls and reflect a small set of inputs the model consistently fails to process.
  • E.5 Obligation Subtype Profiles: Obligation subtype HalTP spans 42.7–88.6%, with a within-bucket SD of 12.4 pp, yet the lowest subtype exceeds the temporal category mean of 29.0–35.1%.Post-Termination Services is highest at 88.6%, while the detailed subtype profile shows substantial intra-bucket heterogeneity.
  • E.6 Debate Pipeline Overhead: In 4,920 clause-level decisions, debate averages 1.12 rounds, with 87.9% finishing in R1 and 12.8% of detections changed.Skeptic–Supporter consensus is 99.94%, with 90.6% stability and 9.4% partial outcomes.
  • E.6 Debate Pipeline Overhead: Per-type flip rates are 4.2% factual, 9.9% temporal, 13.8% numeric, and 14.2% obligation, while mean rounds span only 1.10–1.20.The calibrated benefit therefore comes from which clauses are flipped rather than extended deliberation.
Loading 2606.18021v1…