🤖 machine learning

Evaluating Assurance Cases as Text-Attributed Graphs for Structure and Provenance Analysis

This paper proposes a graph diagnostic framework that models assurance cases as text-attributed graphs, demonstrating that graph neural networks effectively perform link prediction and distinguish between human-authored and LLM-generated cases by identifying distinct hierarchical linking patterns, while also revealing limitations in current explanation methods regarding argument structure faithfulness.

Original authors: Fariz Ikhwantri, Dusica Marijan

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fariz Ikhwantri, Dusica Marijan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a safety inspector for a new airplane. Before the plane can fly, the engineers must hand you a massive "Safety Argument Book." This book doesn't just say "The plane is safe"; it builds a logical tower of reasons, evidence, and claims to prove it. In the industry, this is called an Assurance Case.

For decades, humans have written these books. But now, we have a new tool: AI (Large Language Models) that can write these books instantly. The problem? AI is great at sounding smart, but it sometimes builds "logical towers" that look nice on the outside but have weak foundations inside. It might miss a crucial link between a claim and the evidence, or it might arrange the arguments in a weird, unnatural pattern.

This paper is like a detective story about how to spot the difference between a Safety Argument Book written by a human expert and one written by an AI.

The Big Idea: Turning Text into a Map

The authors realized that reading these books like a novel isn't enough. You need to see the structure.

They took these text-heavy documents and turned them into maps (graphs).

Nodes (The Cities): Each "Claim," "Strategy," or "Piece of Evidence" is a city on the map.
Edges (The Roads): The lines connecting them show how one idea supports another.

In a human-written map, the roads follow a logical, hierarchical pattern (like a family tree). In an AI-written map, the roads might look a bit different—maybe too many roads go straight to the top, or the cities are clustered in strange ways.

The Two Detective Missions

The researchers built a special AI detective (called a Graph Neural Network or GNN) to do two main jobs:

1. The "Missing Link" Detective (Link Prediction)

The Analogy: Imagine you are given a map with some roads erased. Can you guess where the missing roads should go to make the map make sense?
The Goal: The AI tries to learn the "grammar" of safety arguments. If it sees a "Claim" and an "Evidence" node, it should know a road must exist between them.
The Discovery: The AI was very good at this when trained on human maps. But when they tried to teach it using only AI-written maps, it got confused. It learned the AI's "bad habits" and couldn't translate that knowledge back to human maps. However, if they mixed human and AI maps together, the AI detective became super smart, learning the best of both worlds.

2. The "Fake vs. Real" Detective (Graph Classification)

The Analogy: Imagine you are a wine taster. You have two bottles of wine. One is a vintage human-made wine, and the other is a cheap, mass-produced synthetic wine. You can't taste the grapes (the text), but you can look at the shape of the bottle and the label design (the structure).
The Goal: The AI looks at the shape of the argument map. Can it tell, just by looking at the layout, "Ah, this was written by a human" or "This was written by a robot"?
The Discovery: Yes! The AI was incredibly accurate (about 94% correct). It found that AI-generated maps have a specific "fingerprint." They tend to be more rigid, with different patterns of how ideas connect compared to the messy, organic flow of human thinking.

The "X-Ray Vision" Problem (Explainability)

Here is the twist in the story. The AI detective is great at saying "This is fake!" but it's not very good at explaining why.

The Analogy: Imagine a security guard who stops a suspicious person but can't tell you which specific item in their bag was the problem. The guard just has a "gut feeling."
The Finding: The researchers tried to use "X-ray vision" tools to see which parts of the map the AI was looking at. They found that the AI was often relying on the words (the content of the cities) rather than the roads (the connections between them). It was like the guard stopping someone because of their shoes, not because of the illegal item in their pocket. This means we still need humans to double-check the AI's reasoning.

Why Does This Matter?

As we start using AI to write safety reports for things like self-driving cars, medical devices, and nuclear plants, we can't just trust the output blindly.

Safety First: We need to know if the AI missed a critical safety link.
Bias Detection: We need to know if the AI is "hallucinating" or following a lazy pattern that humans wouldn't.
Better Training: The study shows that if we mix human examples with AI examples, we can train better AI tools that understand the true logic of safety.

The Takeaway

This paper is a warning and a guide. It says: "AI can write safety arguments, but it writes them with a different 'handwriting' than humans. We have built a new kind of microscope (Graph Neural Networks) that can spot these differences instantly. However, we still need to teach the microscope to explain its findings clearly so we can trust it with our lives."

1. Problem Statement

Assurance cases are structured arguments used in regulated industries (e.g., aviation, AI safety) to justify that a system meets specific requirements, supported by evidence. Traditionally, these are manually constructed, making them labor-intensive and difficult to scale.

The Challenge: The rise of Large Language Models (LLMs) for automating assurance case generation introduces risks regarding trust, traceability, and bias. LLM-generated arguments may appear structurally valid but often lack logical coherence, omit critical evidential links, or exhibit different hierarchical patterns compared to human reasoning.
The Gap: Existing evaluation methods rely heavily on text similarity or natural language inference, failing to capture the complex inter-dependencies and structural topology inherent in assurance arguments. There is a need for a framework that can analyze the structure and provenance (human vs. AI) of these arguments to detect bias and ensure reliability.

2. Methodology

The authors propose a Graph Diagnostic Framework that models assurance cases as Text-Attributed Graphs (TAGs). In this representation:

Nodes: Represent argument elements (e.g., Claims, Strategies, Evidence, Contexts) with text attributes encoded via pre-trained language models (e.g., BERT).
Edges: Represent the logical relationships (links) between these elements.

The framework addresses two primary tasks using Graph Neural Networks (GNNs):

A. Link Prediction

Goal: To learn and infer missing or implicit connections between argument elements.
Approach: The model predicts the probability of an edge existing between two nodes ( $v_i, v_j$ ) based on learned node embeddings.
Objective: To quantify the model's ability to capture structural and semantic dependencies across different assurance case formats (Tree structures vs. Goal Structuring Notation - GSN).

B. Graph Classification (Provenance Detection)

Goal: To distinguish between human-authored and LLM-generated assurance cases.
Approach: The model aggregates node embeddings into a global graph representation (via pooling) and classifies the entire graph.
Objective: To detect structural and stylistic biases inherent in synthetic data, serving as a diagnostic tool to flag potential hallucinations or generative biases.

C. Explainability Analysis

The study utilizes GNNExplainer to analyze model behavior, measuring fidelity (how well the explanation matches the model's decision) and identifying which nodes/edges drive the classification.

3. Key Contributions

Curated Dataset: The authors compiled a publicly available dataset of assurance cases represented as graphs. It includes:
- Human-authored cases: From safety trees and GSN formats.
- LLM-generated cases: Synthesized using GPT-4o.
- This dataset supports both link prediction and provenance analysis across different structural paradigms.
Empirical Evaluation Framework: A comprehensive study evaluating GNNs (GCN, GAT, GraphSAGE, UniGraph) and Decoder LLMs (Llama, Qwen) on two complementary tasks: structural inference (link prediction) and bias detection (graph classification).
Analysis of Structural Bias: The paper provides evidence that LLM-generated assurance cases exhibit distinct hierarchical linking patterns and node distributions compared to human-authored ones, which can be detected via graph topology.
Explainability Gap Identification: The study reveals that while GNNs achieve high accuracy in classification, existing explanation methods show only moderate faithfulness, indicating a gap between predicted reasoning and the true argument structure.

4. Key Results

Link Prediction (RQ1)

Human-to-Human: GNNs trained on human data perform strongly (ROC-AUC $\approx$ 0.76–0.85), significantly outperforming text-only LLM baselines. This confirms that structural information is critical for link inference.
LLM-to-Human (Cross-Provenance): Models trained only on LLM-generated data perform poorly when tested on human data (significant drop in AUC/F1). This indicates a distributional mismatch and that LLMs fail to learn the true hierarchical reasoning of humans.
Semi-Supervised Learning: Combining human and LLM data for training substantially improves generalization. UniGraph (a graph foundation model) achieved the best results (ROC-AUC $\approx$ 0.868), suggesting that diverse training sources mitigate structural gaps.
Depth Analysis: Shallow models (1-layer) generally outperformed deeper variants, suggesting that link inference in assurance cases relies on local neighborhood information rather than deep global propagation.

Graph Classification (RQ2)

Detection Accuracy: GNNs effectively distinguish between human and LLM cases with high accuracy (F1 $\approx$ 0.94–0.98).
Bias Signals: The models detect that LLM-generated cases have irregularities, such as disproportionate reliance on high-level "Goal" and "Strategy" nodes and inconsistent "Context" integration.
Text-Only Failure: Decoder LLMs (using In-Context Learning) failed to classify provenance accurately (F1 < 0.2), proving that structural topology is the key differentiator, not just text content.

Explainability (RQ2b)

Faithfulness: While models are accurate, their explanations are not always faithful.
- GCN showed balanced attribution.
- GraphSAGE exhibited polarized fidelity (relying on very few nodes).
- Edge-level fidelity was generally low, suggesting models rely more on node features (semantic content) than edge structures for classification, despite the task being structural.

5. Significance and Implications

Safety & Certification: The framework offers a practical method to audit AI-generated safety documentation. It can flag LLM-generated arguments for closer human inspection before certification.
Beyond Text Similarity: The study demonstrates that analyzing the topology of arguments is superior to text-based similarity for detecting AI hallucinations and structural bias in safety-critical domains.
Data Efficiency: The success of UniGraph and GraphPrompt highlights that graph foundation models can learn complex structural semantics with limited data, addressing the scarcity of high-quality human-authored assurance cases.
Future Directions: The findings suggest a path toward iterative refinement of LLMs, where graph classifiers could serve as discriminators in adversarial training loops to improve the structural quality of generated assurance cases.

In conclusion, the paper establishes that assurance cases are fundamentally graph problems. By leveraging GNNs, researchers can not only predict missing links but also detect the "fingerprint" of AI generation, providing a crucial layer of validation for automated safety engineering.