Module validate 

Native validation harness: the agent’s reward function.

Builds the scheduler, attaches it on the host (as root), drives a workload, extracts a single metric, always tears the scheduler down (verifying the kernel returns to disabled), and returns a Verdict. Runs in-process; the equivalent logic previously lived in a separate run_validation.py.

Stages mirror the old harness: spec / build / preflight / attach / metric / complete. A non-complete verdict means the candidate did not build, attach, or produce a measurable result this round.

Structs

Measurement 🔒

One scheduler+workload run.

Scheduler 🔒

Launch/teardown of the scheduler with hard safety guarantees.

TraceRecorder 🔒

TraceStats 🔒

Verdict

Constants

SCX_STATE_PATH 🔒

Functions

binary_help 🔒

Capture the scheduler binary’s --help text (stdout+stderr) so we can probe which optional flags it supports before launching it. Passing an unknown argument makes the scheduler exit before attaching, so flags that are not universal must be gated on this. Empty string on any failure.

binary_path 🔒

cargo maps release/dev to target/{release,debug}; named profiles land in target/.

cargo_build 🔒

cargo build the candidate. Returns (ok, combined stdout+stderr).

cpu_event_rate_summary 🔒

event_counts_json 🔒

event_rates_json 🔒

exec_scheduler_foreground

Replace this process with the scheduler in the current terminal.

extract_metric 🔒

Parse the workload command output as the single numeric metric value. Any extraction or aggregation belongs in the command itself.

fail 🔒

Build a non-complete verdict (build/preflight/attach/metric/spec stages).

measure_once 🔒

median 🔒

perf_available 🔒

perf_mmap_pages_per_cpu 🔒

Per-CPU mmap pages for perf record -m, derived from a total byte budget split across the online CPUs. This controls perf’s live mmap buffer; the output file is separately capped with perf record --max-size.

pstdev 🔒

Population standard deviation (matches Python statistics.pstdev).

rate_value 🔒

read_lossy 🔒

round_trace_float 🔒

rounded_option 🔒

run_capture 🔒

Run a shell command to completion or until timeout, capturing combined stdout+stderr to out_path. A fixed-duration workload is expected to be cut off at the timeout; its partial output is still returned.

run_validation

Build the candidate, attach it, drive the workload, and return a verdict.

scx_is_enabled 🔒

scx_state 🔒

Return the sched_ext root state (“disabled”/“enabled”/…) or None.

sleep_interruptible 🔒

Sleep up to duration, returning true if interrupted before the deadline.

sorted_counts 🔒

summarize_trace 🔒

tail 🔒

trace_cpu 🔒

trace_event 🔒

trace_stats_json 🔒

trace_window_sec 🔒

update_trace_stats 🔒

wait_for_state 🔒

Poll the sched_ext state until it equals want or the timeout elapses.