# Post-first-fault observation — 2026-11-18T05:09–05:16Z

After the first injected fault (node-1 killed 05:03:37Z, restarted 05:04:34Z; 5 expected check
errors, 0 divergences), the soak recorded four consecutive **unexpected** check errors:

| Time (UTC) | Error |
|---|---|
| 05:09:28 | `ReadTimeout` |
| 05:09:47 | `ReadTimeout` |
| 05:10:06 | `ReadTimeout` |
| 05:10:27 | `ReadTimeout` |

`max_unexpected_check_streak = 4`; the soak's verdict limit is 5. The window coincided with a
view change. By 05:13Z the cluster was stable at **view 207**, all four replicas at the same
height, and the gateway reported **HEALTHY**. No divergence was recorded.

To reduce host contention, the long-running `cain_mr_restore_validate.py --publish` process
(PID 2539218, ~31% CPU, 51 min) was reniced to 19 and `ionice -c3`; it was **not stopped**. Host
load at the time was ~38 on 8 CPUs, mostly other agent/gateway workloads. See
`../logs/stabilization.txt`.

This is a disclosed risk, not a failure: the run remains **RUNNING**, and the final verdict comes
from the published verifier at the end of the 48 h (~2026-11-20T04:43Z).
