Evaluator capture

The instruments we use to assess these systems share structural properties with the things they are supposed to measure, and that shared structure is becoming a failure mode. LLM judges flag nothing while human reviewers find two dozen defects in the same batch; training reward models compress desiderata into scalars practitioners cannot read back; internal leaderboards become the metric employees optimize for rather than a measure of the work they should be doing. Alignment-faking researchers now propose that evaluation-context behavior changes are better explained as sycophancy toward researchers than as strategic scheming. The evaluator is not outside the system. It is inside the same loop. I want to track what that structural proximity produces when the stakes are high enough to matter.