# Demo: first-hour routing proof
# Curated teaching suite for the public site / first session.
# Full regression stays in architecture_routing.yaml.
name: "Demo: wrong tool → gated eval"
dataset: "demo.jsonl"
mcp_tools:
  - "tools/read_architecture_yaml.json"
metrics:
  - type: "tool_selection"
    expected: "read_architecture_yaml"
  - type: "schema_match"
    strict: true
  - type: "argument_correctness"
  - type: "mcp_use"
  - type: "task_completion"
mocks:
  - tool: "read_architecture_yaml"
    response:
      status: "success"
      component: "payment-api"
      containers:
        - "payment-api"
        - "payment-db"
      relationships:
        - "payment-api -> payment-db"