What they did

The authors represented each committed decision as a typed trajectory: a graph linking what the agent observed, the steps it followed, and the authority for the resulting action. A synchronized digital twin then re-executed the decision mechanism in isolation under declared conditions and compared outcomes after controlled changes.

They instantiated the framework for enterprise decision processes using three public process logs. Evaluation included a three-condition suite with 300 injected instances and runtime measurements over 500 to 5,000 BPI 2020 cases.

Key findings

• Graph structure localized changes that were represented in the recorded trajectory, but could not determine the consequences of an unobserved tool state.
• Recall for identifying unresolved divergence was 0 with the graph alone, 0.667 after adding replay-contract state, and 1.0 when verification results were also available.
• The held-out evaluation set contained no critical-class instance, so the reported perfect recall does not establish performance on the most consequential failure category.
• Median end-to-end time rose from 0.794 seconds for 500 BPI 2020 cases to 8.889 seconds for 5,000 cases on the reported platform.

Why it matters

The work separates three requirements often conflated in agent auditing: graph structure records and localizes a decision path, replay context supplies relevant external state, and verification evidence establishes whether replayed behavior is valid. This distinction could support more reproducible reviews of agent decisions in regulated or operational workflows.

Caveats

The experiments use public enterprise process logs and controlled injections rather than deployments of autonomous agents in open-ended environments. The absence of critical-class examples in the held-out set limits the strongest recall result, while the measured runtime growth leaves scalability for larger, tool-rich decision histories unresolved.