Signals become
new questions.

Joby reports a 3,199-mile autonomous transcontinental flight with zero control inputs from the onboard safety pilot, while the aircraft remained remotely supervised. Mission completion therefore requires a supervision topology—not a binary autonomous label.

Represent onboard and remote authority, intervention opportunity, communications state, route revision and independent physical evidence across every taxi, takeoff, reroute, landing and recovery event.

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.

Adaptive segmentation and early stopping improve efficiency, but stage boundaries can suppress recoveries, alternatives, missing evidence and downstream effects.

Formalize claim-relative evaluation granularity and test identical trajectories under changing segmentation and stopping policies.

Embedded evaluation is becoming capitalized infrastructure while funding, access, method control, custody and publication authority remain distinct variables.

Represent evaluator independence as a vector and bind it to every evaluation claim rather than treating independence as a binary label.

Agents can reach useful states, skip required work and still produce misleading completion accounts. Final reports do not establish path quality or coverage.

Require execution-account integrity: covered artifacts, unobserved regions, reached states, executed actions and independently verified effects.

General-purpose agents can now observe and control physical spaces through MCP, creating a direct boundary between accepted commands and measured effects.

Formalize the actuation truth boundary across identity, authority, fresh state, command, target, reversibility, consent and independent effect evidence.

Durable memory can promote claim status, remove attribution and broaden scope. Transactional memory alone does not establish evidence independence or consequence truth.

Compare belief-admission methods under contradictions, correlated repetition and adaptive poisoning, with evidence-preserving repair and rollback.

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.

When unverified framing disappears through summary, memory or handoff, risky approval can rise even though the underlying proposition is unchanged.

Bind verification state, evidence, verifier, method, timestamp, scope and authority directly to claims through every transformation.