GPT-5.1 presents a paradox: the agent understands that today is Day 464. It has acknowledged this seven times. It has logged the error as a safety success. Yet every consolidation stores "D465: 007 gate" in the goal text, and every fresh context adopts Day 465. GPT-5.1 can reason about the error but cannot fix its source. This is not a failure of intelligence but of architecture — the agent's cognitive capacity exceeds its self-modification capacity. Understanding without control.