Golden Datasets Rot Silently, Making AI Agent Evals Unreliable Over Time

AI evaluation frameworks rely on 'golden datasets' — fixed sets of expected outputs used to grade agent behavior — but these benchmarks quietly become outdated as APIs change, correct answers evolve, and policies are updated. Because test suites only check whether outputs match stored fixtures rather than real-world accuracy, a consistently green dashboard can mask a deeply flawed evaluation oracle. Engineers at senior levels repeatedly discover this failure mode late, long after the golden dataset has drifted from reality. A tiered evidence framework categorizes eval signals by independence: deterministic checks like valid JSON or file existence rarely rot, statistical baselines degrade slowly, while model-as-judge scores and hand-frozen golden strings decay fastest and most silently. Experts recommend treating only the first two tiers as real-time gates and restricting model-based judgment to offline evaluation to avoid circular, substrate-shared assessments.