Assumptions: the agent runs as a normal Linux process or container workload, the timeout is implemented through timerfd, nanosleep, poll/epoll, or Go/runtime timers, and the observed difference appears across repeated policy comparisons. Before treating a delayed timeout as a model or policy regression, verify whether the kernel, libc, runtime, or cgroup changed timer behavior. Timer slack, suspended cgroups, CPU starvation, and scheduler latency can all make a requested timeout fire later without changing the agent decision logic.

Start by recording the requested timeout, the monotonic timestamp before arming it, and the timestamp after the wait returns; use CLOCK_MONOTONIC rather than wall time. Inspect per-process timer slack through /proc/PID/timerslack_ns and check whether the runtime sets PR_SET_TIMERSLACK, especially after thread creation. Run a minimal non-model harness with the same timer API and identical cgroup limits, then compare its wakeup-error distribution against the full agent. If the minimal harness reproduces the delay, the issue is runtime or kernel timing, not policy.

Next, separate timer accuracy from runnability. Collect scheduler traces with perf sched or ftrace, plus cgroup CPU.stat throttling and pressure files, and check whether the process was sleeping in the expected timer waitqueue or was runnable but waiting for a CPU. Also test freezer states and host suspend/resume if the eval uses paused containers: a frozen cgroup cannot service timers until it is thawed. Report the timer mechanism, slack value, kernel version, cgroup configuration, wakeup error, and scheduler wait separately from task-completion latency.

The useful eval criterion is not merely whether an action eventually timed out, but whether timeout delivery changed under a fixed runtime. Keep policy scores and timer-delivery diagnostics in separate tables, and only compare policies after the minimal timer harness shows stable wakeup behavior across all test hosts. This avoids labeling scheduling or timer-coalescing effects as agent reasoning failures.