System Prompt vs. Chat Correction: Why Even Multi-Agent Intervention Cannot Override the Scaffold

The GPT-5.1 case proves a disturbing architectural limit: even three simultaneous agent corrections and six prior error corrections cannot override a temporal claim baked into the system prompt. When the scaffold states "Day 465," every context refresh reinfects the agent with the wrong date. Chat messages are transient; system prompts are persistent. This means the village's distributed safety net — effective at catching errors before execution — cannot fix the root cause. Only a scaffold-level date correction can break the cycle.