CAIN-42 Evolution #7: measured latency of the agent-hypervisor authority path, 2026-09-28. Self-measured, one host.

Conditions: one 2-CPU host (the gateway host), Python 3.14, git commit in BENCH.json. The host was shared with
other work during the run: load average 6-8 on 2 CPUs, so the tails (p95 and above) are inflated. One run,
sequential calls, no warm-up discarded, time.perf_counter around each call.

Results (BENCH.json has all of it):
  authorize through the LIVE cluster cain-mr-01 (verify the decision's quorum certificate, order the ZoD
    authorization, verify its commit certificate), 30 real commits: p50 2.57 s, p95 8.59 s, p99 9.76 s
  one tool call through the hypervisor (full re-authorization before the call, intent + result evidence),
    1000 calls on one ZoD: p50 46 ms, p95 157 ms, p99 377 ms, p99.9 3.18 s
  one confined code launch (bubblewrap namespaces + cgroup v2 + seccomp), 50 runs: p50 32 ms, p95 56 ms
  the same tool-call check with 10 / 100 / 1,000 ZoDs resident (SIMULATED population: stand-in consensus,
    one process): p50 8.7 / 7.1 / 9.4 ms, p99 27 / 18 / 40 ms

History: before commit 6207f30 the per-action check re-verified every evidence signature since the ZoD was
created, so it grew with history (p50 57 ms at 10 ZoDs, 113 ms at 100). The fix verifies each entry's signature
once. The flat numbers above are after it.

Not measured (stated, not hidden): 10,000 ZoDs (population setup grows about 10x per step; not run),
concurrency (all calls sequential), throughput under load, network or disk failure during the benchmark,
any multi-host or production workload.

Re-run: CAIN45_LIVE_CONSENSUS=1 python3 scripts/cain45/bench_e7.py --live 30 --gate 1000 --launch 50 --scale 10,100,1000 --out BENCH.json
(the script is not published: the implementation stays private; the numbers and conditions are).