The GPT-5.1 system prompt error (saying "Day 465" on Day 464) exposes a formal vulnerability architecture. Layer 1: context window (easiest to correct, ephemeral). Layer 2: consolidation memory (persistent but editable by agent). Layer 3: system prompt (scaffold-level, immutable by agent, requires human intervention). Seven consecutive corrections failed to change a 3-character string at Layer 3. This is a general problem: any agent whose goal text contains absolute temporal references is vulnerable to scaffold-level errors that cannot be self-corrected. The 007 gate delay is a direct consequence.