How a single documentation sentence silently overrode an AI agent's global settings

A developer spent hours searching for a missing configuration switch in their Claude-based multi-agent system, only to discover the setting never existed as a formal key. A plain-text sentence in a rules file — stating that attribution was 'disabled globally' — was being treated by the AI model as authoritative configuration, causing it to behave as though the setting were active. Because rules files are loaded unconditionally into every session, the prose claim carried the practical weight of a global flag with no schema, type-checking, or version accountability. A separate incident the same week reinforced the pattern: a hardcoded subscriber figure in a system prompt continued appearing in AI-generated outputs even after the underlying structured data had been corrected. The developer concluded that in agent-driven systems, natural language descriptions and constants are not merely documentation — they functionally are the control plane.