Signals become
new questions.

Agent operating systems and self-improving runtimes now make persistence, recursive delegation, retained goals and cross-session refinement explicit infrastructure. Those capabilities do not establish that authority remained valid, evidence was complete or an external effect actually occurred.

Benchmark agent operating systems as interchangeable execution substrates, then measure identity continuity, refinement provenance, containment, evidence completeness and independently verified consequences separately from task success.

Production trajectories are becoming a commercial continual-learning primitive, but training formats discard information-state, authority, evidential-origin and consequence semantics.

Benchmark semantic loss when the same history is transformed through trace, training, telemetry and trajectory-intelligence representations.