DNative-Twin replays agent decisions, names its blind spot

DNative-Twin records an agentic decision as a typed trajectory: observed state, the path taken, and the authority behind it. Then it replays that trajectory alone, outside the rest of the system.

That matters for teams shipping agents that act without a human checking each step. In a 300-instance test, the system started at zero recall on decisions it couldn't fully explain.

The paper names its own limit. Graph structure can localize which parts of a decision changed, but it can't determine what a tool did in a state nobody observed.

Adding replay-contract state raised recall on those unresolved cases to 0.667. Adding verification results on top of that pushed it to 1.0.

Verification results account for most of the gain. Without them, DNative-Twin misses roughly one in three of the decisions it can't fully explain.