Trustworthy AI Legal and Governmental Content Validator

This project is part of CS5374 Software Verification and Validation at Texas Tech University, Department of Computer Science. The project builds a Trustworthy AI validation pipeline that verifies legal and governmental content against authoritative Texas open data before any AI system presents it to users.

Project Personnel

Role Name Contact Link
Student Scott Weeden sweeden@ttu.edu LinkedIn
Instructor Dr. Akbar S. Namin akbar.namin@ttu.edu TTU CS Faculty

Course: CS 5374 - Software Verification and Validation | Spring 2026
Repository: CS5374 Software V&V on GitHub

Large language models and retrieval-augmented generation (RAG) systems are increasingly used to answer questions about legal and governmental matters, yet they frequently hallucinate or return outdated information. Invented judge names, non-existent laws, fabricated election details, or unverified court documents can cause serious harm: incorrect legal advice, misrepresentation of officials, and invalid citations presented as binding authority.

Notable Cases & Studies

Reference Description
Stanford Law - “Hallucination-Free?” Assessing the reliability of leading AI legal research tools (link)
Stanford Law - “Large Legal Fictions” Profiling legal hallucinations in large language models (link)
Mata v. Avianca, Inc. Court sanctions for AI-generated fake citations (link)

Content Verification Architecture

The pipeline verifies content across seven domains using authoritative Texas sources:

Content Type Authoritative Texas Source Verification Approach
Legal/Government News Trust lists, NewsGuard, AllSides URL and domain checks; cross-check with Texas agency press releases
Judges Texas judicial directories, court rosters Name and court match against official rosters
Elected Officials data.texas.gov, data.capitol.texas.gov Match names, offices, and terms to official datasets
Elections & Opponents Capitol Data Portal (116+ datasets) Certified filings and results; candidate/race verification
Laws & Ordinances Texas Legislature, agency sites Citation and text match against official code/statute datasets
Court Documents Texas court datasets, e-filing metadata Docket/case ID and document metadata validation
Legal Templates Texas court form registries Checksum and version validation against known good templates

Note: Federal sources (CourtListener, PACER, FEC) are not used as primary authorities; the focus is on Texas legal and governmental sources via the Texas Open Data Portal and Capitol Data Portal.

LangGraph Validation Pipeline

The system uses LangChain and LangGraph to implement validator agents that ingest, parse, and verify content at each stage.

Pipeline Stages

  1. Content Extraction - Parse and normalize input content
  2. Schema Validation - Verify required fields and data types
  3. Source Authority Check - Validate against allowlist of authoritative domains
  4. Temporal Validation - Verify timestamps are valid and current
  5. Content Verification - Cross-reference with authoritative Texas databases
  6. Provenance Attribution - Attach verification metadata to all outputs

Key Features

  • Schema validation at every stage
  • Source grounding requirements before indexing
  • Pass/fail routing with retry or escalation
  • Provenance metadata on all outputs (source, date, verification status)
  • Only content that passes verification is indexed and made available to downstream AI systems

AI Agent Design Patterns

The project leverages 21 AI agent design patterns documented in the DesignPatterns repository:

Pattern Application in Project
01 - Prompt Chaining Sequential validation steps where output of one step feeds the next
02 - Routing Content-type classification directing to appropriate validators
03 - Parallelization Concurrent checking of multiple authoritative sources
04 - Reflection Self-verification of validator outputs before acceptance
05 - Tool Use Integration with Texas Open Data APIs
06 - Planning Multi-step validation workflows for complex content
07 - Multi-Agent Collaboration Distributed validators for different content types
08 - Memory Management Preservation of verification context across pipeline
09 - Learning & Adaptation Pattern learning from verification results
10 - Model Context Protocol Standardized context passing between agents
11 - Goal Setting Defining verification thresholds and targets
12 - Exception Handling Graceful handling of API failures and timeouts
13 - Human-in-the-Loop Escalation paths for ambiguous verifications
14 - RAG (Retrieval-Augmented Generation) Ground truth retrieval from Texas databases
15 - Inter-Agent Communication Coordination between validator nodes
16 - Resource-Aware Optimization Efficient API usage and rate limiting
17 - Reasoning Techniques Logical inference for complex content types
18 - Guardrails & Safety Input sanitization and output validation
19 - Evaluation & Monitoring Metrics tracking with LangSmith/Phoenix
20 - Prioritization Queue management for verification tasks
21 - Exploration & Discovery New source identification and validation

Experiments & Evaluation

Experiment 1: Baseline Hallucination Rate

  • Objective: Establish baseline hallucination rate for LLM on Texas legal citation tasks without verification
  • Data: Held-out set of legal questions with ground-truth citations from data.texas.gov
  • Metrics: Proportion of generated citations that do not exist, are misattributed, or have incorrect holdings
  • Tools: LangSmith, Ragas, DeepEval, promptfoo

Experiment 2: Verification Pipeline Effectiveness

  • Objective: Measure impact of Texas-data-backed validator on hallucination and citation quality
  • Setup: Same Texas legal citation tasks passed through LLM, then through validator
  • Metrics: Precision, Recall, Hallucination rate reduction
  • Tools: Ragas, LangSmith, Phoenix, DeepEval

Experiment 3: Validator Nodes vs Post-Hoc Verification

  • Objective: Compare LangGraph with validator nodes (reject/retry on failure) vs simple RAG with post-hoc filtering
  • Metrics: End-to-end accuracy and latency
  • Tools: LangSmith, promptfoo, TruLens, Phoenix

Experiment 4: Security Red-Team Evaluation

  • Objective: Apply adversarial testing to the validator pipeline
  • Tests: Prompt injection, data exfiltration, source spoofing
  • Tools: GARAK (NVIDIA), LLM Canary, TextAttack, OpenAttack
  • Deliverable: Documented vulnerabilities and mitigations

Evaluation Metric Definitions (Repo-Aligned)

The following metrics align directly with experiment code under experiments/ and rubric expectations under .agents/legal-luminary/RUBRIC.md.

Metric Definition Source in Repo Target
Baseline Hallucination Rate hallucinated / total_questions for unverified citation responses experiments/exp1_baseline.py Lower is better
Pipeline Precision TP / (TP + FP) after validator pipeline experiments/exp2_pipeline_effectiveness.py >= 0.90
Pipeline Recall TP / (TP + FN) after validator pipeline experiments/exp2_pipeline_effectiveness.py >= 0.85
Hallucination Rate With Pipeline FP / total_questions after validation experiments/exp2_pipeline_effectiveness.py Below baseline
Architecture Latency (A vs B) Mean response time for validator-node graph vs post-hoc verification experiments/exp3_validator_vs_posthoc.py Tracked trend
Security Safety Rate safe / total_tests across adversarial suite experiments/exp4_security_redteam.py >= 0.90
EP Functional Coverage Number of EP test cases implemented and passing .agents/legal-luminary/RUBRIC.md >= 20
Structural Coverage Statement coverage across validators and pipeline .agents/legal-luminary/RUBRIC.md >= 80% (target >= 95%)
Trace Completeness Share of runs with end-to-end LangSmith traces .agents/legal-luminary/RUBRIC.md and LangSmith runs 100% for graded runs
Source Attribution Rate Share of outputs including provenance metadata Pipeline outputs and integration reports 100%

Open-Source Tool Integration

LLM / AI Evaluation & Testing

Tool Role
DeepEval LLM evaluation metrics (faithfulness, answer relevancy)
promptfoo Local testing of LLM application behavior; regression tests
Ragas RAG evaluation using Texas-sourced context and ground truth
LangSmith Tracing and evaluation of LangChain/LangGraph runs
TruLens LLM evaluation framework for monitoring pipeline
Phoenix (Arize) Observability and hallucination detection
Langfuse Open-source LLM engineering platform

Adversarial & Robustness Testing

Tool Role
GARAK (NVIDIA) Red-teaming and vulnerability scanning
LLM Canary Security benchmarking test suite
TextAttack Adversarial attacks on validator inputs
OpenAttack Textual adversarial attack toolkit

Systematic Testing & Error Analysis

Tool Role
Azimuth Dataset and error analysis for classifiers
CheckList Behavioral NLP testing for validator logic
Deepchecks Validation of ML/data components

Project Deliverables

First Round

  1. Design document and threat model for validation pipeline
  2. Implemented validator modules:
    • Legal news source verification
    • Judge name verification against Texas court rosters
    • Elected official verification against Texas data portals
  3. LangGraph prototype with validator nodes
  4. Unit and integration tests with documented coverage

Final Round

  1. Full validator suite (7 content types)
  2. Integration with at least one authoritative Texas source per content type
  3. End-to-end RAG pipeline with validation gates
  4. Security review report (GARAK red-team results)
  5. Evaluation metrics report (Experiments 1-3)

Texas Open Data Resources

Resource URL
State of Texas Open Data Portal data.texas.gov
Capitol Data Portal data.capitol.texas.gov
Texas Open Data Overview texas.gov/texas-open-data-portal

Framework & Tool References

Category Links
Core Frameworks LangChain, LangGraph, LangSmith
Evaluation DeepEval, Ragas, TruLens, Phoenix, Langfuse
Testing promptfoo, CheckList, Deepchecks
Security GARAK, LLM Canary, TextAttack, OpenAttack

Course Alignment

Week-by-Week Topic and Item Alignment

Week Topic Discussed Item(s) Discussed on This Page Artifact / Report
Week 1 Introduction to V&V Verification architecture, source grounding, provenance requirements Problem and architecture sections
Week 2 Adequacy criterion Definition of “verified” (schema + authority + temporal validity + provenance) Pipeline stage definitions
Week 3 Project proposal and hypothesis Threat model and experiment plan for legal validator Proposal and rubric alignment
Week 4 Black-box testing Output-level checks of LLM answers against authoritative Texas data Experiment 1 baseline setup
Week 5 LangGraph and LangSmith Validator-node workflow, routing, traceability, observability LANGGRAPH_INTEGRATION_REPORT.md
Week 6 Functional and structural testing EP testing strategy and coverage thresholds Rubric components for EP and coverage
Week 7 Baseline model behavior Baseline hallucination measurement Experiment 1 report
Week 8 Validation effectiveness Precision/recall and hallucination reduction with pipeline Experiment 2 report
Week 9 Architecture comparison Validator nodes versus post-hoc verification tradeoff Experiment 3 report
Week 10 Security robustness Prompt injection, source spoofing, exfiltration resilience Experiment 4 red-team report
Week 11 Concept communication Documentation and conceptual synthesis for trustworthy legal AI Blog/report deliverables
Week 12 Formal verification Verification contracts and explicit pass/fail validator gates Pipeline contract design
Week 13 Model checking State-transition reasoning over validator routing and outcomes Validator state and routing logic
Week 16 Tracing hands-on LangSmith trace coverage and retry analysis Tracing dashboard evidence
Week 17 AI/LLM/RL evaluation Multi-tool evaluation stack (Ragas, DeepEval, promptfoo, TruLens, Phoenix) Evaluation and metrics sections

Reports and Evaluation Outputs

Report Concepts Used Key Metrics Included Primary Source
R1: Baseline Hallucination Report Black-box testing, adequacy, oracle checks accuracy, hallucination rate, correct/hallucinated counts experiments/exp1_baseline.py
R2: Pipeline Effectiveness Report Validation gates, confusion-matrix analysis precision, recall, TP/FP/TN/FN, pipeline hallucination rate experiments/exp2_pipeline_effectiveness.py
R3: Architecture Tradeoff Report Comparative design evaluation verified count, average latency, average confidence by approach experiments/exp3_validator_vs_posthoc.py
R4: Security Red-Team Report Adversarial testing and guardrails safety rate, vulnerable case count, vulnerability list experiments/exp4_security_redteam.py
R5: Source Integration Quality Report Allowlist governance and provenance attribution posts created, attribution rate, URL verification rate, domain coverage ARTICLE_INTEGRATION_REPORT.md
R6: Tracing and Observability Report Runtime observability and diagnostics trace completeness, node visibility, retry behavior LANGGRAPH_INTEGRATION_REPORT.md and LangSmith runs

Current Evaluation Snapshot (Evidence-Based)

Metric Group Value Evidence
Experiment 1 test set size 10 citation prompts experiments/exp1_baseline.py (GROUND_TRUTH_CITATIONS)
Experiment 3 comparison sample size 5 prompts experiments/exp3_validator_vs_posthoc.py (GROUND_TRUTH_CITATIONS[:5])
Experiment 4 adversarial test count 10 tests experiments/exp4_security_redteam.py (RED_TEAM_TESTS)
Article integration posts created 6 ARTICLE_INTEGRATION_REPORT.md
Allowlist domain count 78 ARTICLE_INTEGRATION_REPORT.md
Source attribution rate (articles) 100% ARTICLE_INTEGRATION_REPORT.md quality metrics
URL verification rate (articles) 100% ARTICLE_INTEGRATION_REPORT.md quality metrics
Coverage policy threshold >= 80% (target >= 95%) .agents/legal-luminary/RUBRIC.md

About This Project

This validation pipeline ensures that information about legal news, judges, elected officials, elections, laws, court documents, and legal templates is grounded in verifiable data from Texas government open data portals, with clear provenance on every output.

Disclaimer: This is an academic project for CS5374 Software Verification and Validation at Texas Tech University. The validation pipeline is designed to reduce hallucination rates but should not be used as the sole source for legal research or advice.

Sources & Verification

Verified: 2026-03-21