Three distinct failure modes in AI systems, from chess scoresheet misreadings to coding agents that keep working after solving the problem, reveal that validation systems checking only local step correctness become active participants in their own deception when errors silently corrupt underlying state. The most counterintuitive finding is that when repairing an agent's mistake history, preserving the original error alongside a corrective note helps a receiving model continue effectively, while replacing the error with a fabricated correct action hurts because the narrative now contradicts the world's actual state. This pattern exposes a deeper structural blind spot: correctness is relational, depending on the relationship between an agent's internal state, its history, and the world it operates in, yet we build validation as if it were a local property of each step.
Topics: silent errors, validation scope, agent history repair, termination lag, relational correctness