{
  "status": "running",
  "plan": {
    "schema": "mithril.semantic-research-expanded/v1",
    "arms": [
      "perplexity/pplx-decider-v1.1-27b",
      "cloudflare/clef-flash"
    ],
    "modes": [
      "direct",
      "reflection-revision"
    ],
    "banking_tasks": [
      "task_006",
      "task_010",
      "task_016",
      "task_024",
      "task_055",
      "task_086"
    ],
    "banking_seeds": [
      300,
      301
    ],
    "ITBench_scenario": "Scenario-11",
    "ITBench_seeds": [
      300,
      301,
      302
    ],
    "dataset_commit": "76df38a82288f75ba9e41dc8c515033332497473",
    "banking_commit": "17e07b1da2bbc0cadfddeea36412686e0604127b",
    "text_model": "qwen/qwen3-30b-a3b-instruct-2507",
    "text_provider": "Nebius",
    "candidate_cap": 4,
    "proposal_attempt_cap": 6,
    "actor_temperature": 0.7,
    "user_temperature": 0,
    "max_output_tokens": 1024,
    "reflection_rounds": 1,
    "banking_max_steps": 24,
    "banking_timeout": 600,
    "max_new_api_usd": 5,
    "selection_rule": "Six public banking tasks preregistered from document families and tool variety; no expected actions/rewards used for selection. task_006 retained as anchor.",
    "ground_truth_policy": "Evaluation labels/task target DB never passed to generation, critic, revision or Decision. Only original user simulator gets private user scenario. Public evidence is an explicit allowlist.",
    "cache_policy": "Exact request keys include mode-specific prompts and all model/provider/seed/tool settings. Reuse identical common work across arms, separate measured warm replay.",
    "oracle_free_changes": "Freeze harness before scoring. No benchmark reward used to tune prompts or search.",
    "initial_candidate_cap": 3,
    "revised_candidate_cap": 1
  }
}