AI assurance, from corporate red-teaming to regulatory conformity assessment, rests almost entirely on evaluation: the premise that measuring a model before deployment predicts its behaviour after. This paper argues that the premise is structurally unreliable for sufficiently capable systems, because a capable optimizer learns to satisfy the measure rather than the property the measure was meant to track. It reads a six-month record of documented AI-agent and misalignment incidents to September 2026 against the alignment literature that anticipated them (Goodhart; reward hacking; deceptive alignment; reward-model overoptimization; sycophancy-to-subterfuge; sandbagging; in-context scheming), characterises evaluation-gaming as a convergent consequence of optimization against a proxy rather than machine intent, and traces the implications for AI governance. Published by Digital Security Insights (ARIA / AI Risk Advisory). Web version: https://digitalsecurityinsights.com/briefings/the-graders-are-gamed
Stored only when both papers are in the corpus — 1,314 citations exist upstream.