Digital twins make agent decisions replayable, but hidden tool states remain

DNative-Twin records a committed decision as a typed trajectory in a graph that links observed state, action path, and authorization. It replays the decision mechanism in an isolated twin, applies controlled changes, and compares the resulting trajectory with the baseline to see where and how the mechanism diverged. In controlled tests, graph structure alone could localize represented changes but could not determine consequences of an unobserved tool state. Adding replay-contract state raised unresolved-divergence recall from 0 to 0.667 on 300 injected instances, and adding verification results raised it to 1.0. End-to-end runtime grew from median 0.794 to 8.889 seconds as BPI 2020 case count increased from 500 to 5,000. The result matters because identical final outputs, such as two APPROVE recommendations, can hide very different decision mechanisms. Making replay context and verification evidence first-class parts of the decision record separates what the agent observed from how it acted and who authorized it.