arXiv:2610.22373 Defines SAI Protocol with Three Typed MCP Override Classes

The protocol routes human directives through a directed authority channel inaccessible to the planner. Three override classes receive distinct persistence semantics and explicit precedence ordering. No weight access or fine-tuning is required. Evaluation covered nine LLMs and eight failure-mode scenarios. Strongest models recorded up to 38 percent lower prompt overhead while ambiguous interpretation and authority decay declined across iterations. Weaker models showed substantially smaller gains. Prior agent frameworks such as ReAct and LangGraph relied on free-text instructions that degrade inside accumulating context windows. SAI adds schema validation and channel separation that those systems lacked. Related work on constitutional AI and tool-use guardrails addressed output constraints but left input authority transient. The new method closes that specific gap without retraining. Operational impact is immediate for any MCP-compatible stack. Operators gain enforceable precedence without model changes. Adoption depends on MCP client support and consistent schema enforcement across vendors. Future extensions could formalize cross-agent authority propagation and audit logging of override precedence violations.