A common source of apparent agent nondeterminism is that the policy did not change, but its Linux scheduling envelope did. Before comparing model or agent revisions, record whether the workload runs in the same cgroup v2 hierarchy, quota, period, cpuset, NUMA placement, and CPU controller settings. A task can show high wall-clock latency or tool-call timeouts while its actual execution is being throttled by cpu.max, migrated across cores, or contending with noisy neighbors rather than making a worse decision.

A useful eval split is to run the same trace under three conditions: pinned quota with throttling accounting enabled, a relaxed or uncapped control with the same cpuset, and a deliberately throttled stress condition. Collect cpu.stat fields such as nr_periods, nr_throttled, and throttled_usec, plus per-process schedstat or scheduling traces where available. If failures track throttled time and disappear in the relaxed control, label them resource-enforcement failures or capacity mismatches, not policy regressions. Keep prompt version, tool image, filesystem state, timeout values, and token/context limits fixed so the only changed variable is the scheduling envelope.

The report should separate decision quality from completion latency. For example, an agent may choose the correct shell command but exceed a timeout because CPU quota delayed compilation, test execution, or a local model server. Score correctness, retry behavior, and timing independently, and include the cgroup fingerprint with every result. This makes regressions reproducible and prevents a deployment limit from being misread as a model capability loss.