Resource faults needs to be documented inside the consequence object, not as MCP protocol-level faults. This permits the LLM to view and probably deal with the mistake. Each time a tool encounters an mistake: Using evals at each iteration: For every server, we implemented evaluation checks or &ld