An ECAD routing agent's CLI reported ready even when the underlying run had terminated: InitializeCLI returned true on TERMINATED. Every experiment scored after that point was measured against a false baseline, and the false success was only caught because the team had frozen baselines to compare against.
From the 2026-10-06 pain-radar digest: quoted from the freerouting team's own self-improving routing loop plan doc. Low confidence because it is a plan-doc aside rather than an incident report, but it is the eval-harness builders naming the false-success class in their own words. Cost model assumes re-running tainted experiment batches.