{
  "benchmark": "agenticwire-mcp-vs-rag",
  "sourceDate": "2026-08-16",
  "machine": {
    "cpu": "Apple M1",
    "os": "Darwin 26.3",
    "arch": "arm64",
    "node": "v22.18.0",
    "python": "Python 3.13.5"
  },
  "dataset": {
    "queryCount": 12,
    "documentCount": 8,
    "repetitionsPerQuery": 1000
  },
  "methods": {
    "retrievalOnly": {
      "evidenceRecallPct": 83,
      "structuredExactnessPct": 0,
      "negativeSafetyPct": 0,
      "failureCount": 2,
      "protocolErrorCount": 0,
      "medianLocalLatencyMs": 0.0135
    },
    "mcpReadTool": {
      "evidenceRecallPct": 67,
      "structuredExactnessPct": 100,
      "negativeSafetyPct": 100,
      "failureCount": 4,
      "protocolErrorCount": 5,
      "medianLocalLatencyMs": 0.0005
    },
    "hybridRouter": {
      "evidenceRecallPct": 92,
      "structuredExactnessPct": 100,
      "negativeSafetyPct": 50,
      "failureCount": 1,
      "protocolErrorCount": 0,
      "medianLocalLatencyMs": 0.0009,
      "retrievalRouteCount": 5,
      "mcpRouteCount": 7
    }
  },
  "limitations": [
    "This is not an LLM generation benchmark.",
    "The corpus and query set are intentionally small and hand-authored.",
    "Latency excludes network, embeddings, authorization, and model generation.",
    "The hybrid rule is deterministic and not a learned router."
  ]
}