This repository contains the official implementation of the paper:

Evaluating Assurance Cases as Text-Attributed Graphs for Predicate Structure and Provenance Analysis
EASE 2026 AI

AssureGraph introduces a graph evaluation framework for analysing the semantic and structural patterns of assurance cases. These are structured argument documents used in safety, security, and regulatory compliance.

We model assurance cases as Text-Attributed Argument Graphs (TAGs) and evaluate them using Graph Neural Networks (GNNs) for:

• Link Prediction — identify connections between argument elements
• Graph Classification — distinguishing between human-authored and LLM-generated cases
• Explainability — analysing node/edge importance using GNNExplainer

This repository provides:

• A cleaned and curated public dataset of assurance cases
• Scripts for graph construction, training, and evaluation
• Reproducible experiments for link prediction, provenance classification, and GNN explainability
• Utilities to visualise structural differences between human and LLM-generated cases