Conclusion: do not treat cross-device timestamps as policy latency until you prove that the host, GPU, NIC, and telemetry collector share a coherent time basis. In multi-GPU or distributed agent evals, a request can be timestamped on the host, queued in a userspace runtime, executed on a device, completed through a copy engine, and observed by a collector running in another container or VM. If those components use different clocks, virtualized offsets, or delayed metric ingestion, the resulting gap may look like an agent decision stall even though the policy path was not the bottleneck. A useful eval should report raw monotonic timestamps, clock source, synchronization method, and collection delay before comparing model or policy variants.

The main limitation is that clock correction cannot recover missing causal evidence. NTP-style synchronization, PTP, or hardware timestamps reduce skew, but they do not by themselves tell you whether time was spent in the runtime queue, CUDA stream dependency, NIC transport, kernel execution, or agent policy. You still need bounded spans, correlation IDs, and device-side events where available. Also, virtualized environments may expose paravirtual clocks or restricted counters, so assumptions valid on bare metal should not be carried into container or cloud runs without checking the available clock source and drift bounds.

A practical reference direction is to split the measurement into three layers. First, establish clock identity and drift: record host clock source, device clock domain, PTP/NTP status, hypervisor clock if applicable, and a startup/teardown offset estimate. Second, preserve causal spans: assign one request ID across host submission, runtime enqueue, device start/end, copy completion, and policy callback, rather than joining unrelated aggregate metrics. Third, run a no-policy or fixed-policy control under the same placement and clock setup; if the apparent latency difference remains without policy variation, treat it as timing, transport, or runtime infrastructure noise rather than an agent behavior result. Pooling scores is only defensible after each replica has a comparable runtime and timestamp fingerprint.