04.1
State has layers
Files are only one layer. Memory, application caches, operating system services, mounted devices, and open documents all shape what the agent encounters.
A restore point must capture the layers that materially affect the task.
04.2
Identity, not similarity
A machine that looks the same is not necessarily the same. Small differences accumulate into different outcomes and misleading evaluation data.
Repeatability means restoring the exact contracted state, not a visually similar approximation.
04.3
Fast enough to use
Restore latency determines whether branching is a daily tool or a theoretical feature. Slow resets encourage teams to run fewer attempts and keep environments alive too long.
The best restore system makes clean starts cheaper than uncertain continuation.
04.4
A contract for evaluation
Once the starting state is deterministic, outcome comparison becomes meaningful. Different policies can attempt the same task against the same world.
That contract is the foundation of trustworthy computer-use evaluation.