{
  "benchmark_id": "OTM-Persona-Align",
  "version": "v1.0",
  "run_id": "2026-05-25T10:45:31.000501+00:00",
  "completed_at": "2026-05-25T10:45:31.000501+00:00",
  "agent_version": "v0.2.0-oos-clean",
  "n_samples": 49,
  "seeds": [
    42
  ],
  "metrics": {
    "primary": {
      "name": "response_divergence_cosine",
      "value": 1.0011,
      "ci_low": 0.9939,
      "ci_high": 1.0082,
      "n": 49
    },
    "secondary": [
      {
        "name": "skill_selection_jaccard",
        "value": 0.0466,
        "ci_low": 0.0163,
        "ci_high": 0.0833
      },
      {
        "name": "mean_response_length_execution",
        "value": 677.1
      },
      {
        "name": "mean_response_length_analysis",
        "value": 664.6
      }
    ]
  },
  "baselines": [
    {
      "name": "fixed_neutral_mode",
      "primary_value": 1.0043,
      "primary_ci_low": 0.997,
      "primary_ci_high": 1.0116,
      "jaccard_value": 0.0395,
      "p_vs_agent": 0.699452228257325,
      "description": "Two neutral (midpoint) runs \u2014 measures noise floor. Persona module is working if agent divergence >> baseline."
    }
  ],
  "notes": "Total agent runs: 196; Total LLM cost: $1.77; Cost caps: per-turn=$0.20, total=$4.00; Mean response length: exec=677 chars, anal=665 chars; Wilcoxon p=0.6995 (agent > baseline)",
  "data_filter": "All queries synthetic; no OOS data boundary involved. Fresh agent instance per run (no memory carryover). Embedding model: hash-fallback (non-semantic)."
}