Establish whether governance benefits persist under live execution

Establish whether the improvements in user understanding, control, and trust observed with TaskVM under replayed execution persist when a live model executes participant-authored tasks.

Background

The controlled user study compared TaskVM with a conversational computer-use-agent interface while holding the environment and replayed agent trajectory constant. Under those conditions, TaskVM produced higher perceived recovery control and trust, but the design isolated interface perception from the variability and failures of live model execution.

The paper explicitly leaves unresolved whether those benefits generalize to open-ended tasks authored by users and executed by a live model. Stage 3 explores this setting descriptively, without a conversational control condition or adjudicated per-task success, so it does not resolve the comparative question.

References

Stage 2 held the environment and the replayed agent trajectory constant by design, which isolates perception but leaves open whether these gains persist when a live model, rather than a replay rig, executes tasks the participants themselves invent.

— Task-State Governance: A Virtual Machine for Verifiable and Recoverable GUI Agent Execution  (2610.20100 - Yang et al., 14 Oct 2026) in Section 6, Stage 2: User Study on Task Governance, subsection “User Study Results”