{
  "baseline": {
    "brier": 0.2513,
    "confusion": {
      "fn": 0,
      "fp": 74,
      "tn": 0,
      "tp": 76
    },
    "cost_per_customer": 0.9866666666666667,
    "log_loss": 0.6957513904359482,
    "n": 150,
    "prevalence": 0.5066666666666667,
    "roc_auc": 0.5,
    "threshold": 0.05
  },
  "business": {
    "contribution_winner": 25,
    "conversion_winner": 20,
    "currency": "fictional currency units",
    "inventory_example": {
      "inventory": 10,
      "mean_demand_both": 10,
      "served_deterministic_demand_10": 10,
      "served_half_0_half_20": 5
    },
    "pricing": [
      {
        "conversion_probability": 0.3,
        "expected_contribution_per_opportunity": 3.0,
        "price": 20,
        "unit_cost": 10
      },
      {
        "conversion_probability": 0.24,
        "expected_contribution_per_opportunity": 3.5999999999999996,
        "price": 25,
        "unit_cost": 10
      },
      {
        "conversion_probability": 0.15,
        "expected_contribution_per_opportunity": 3.0,
        "price": 30,
        "unit_cost": 10
      }
    ],
    "randomized_experiment": {
      "arm_sizes": {
        "0": 374,
        "1": 426
      },
      "assignment_probability": 0.5,
      "decision": "inconclusive",
      "delivered_by_arm": {
        "0": 0,
        "1": 385
      },
      "delivery_count": 385,
      "intent_to_treat_contribution_difference": -0.2194270794105151,
      "intention_to_treat": {
        "accounts_per_arm_to_detect_designed_effect": 4912,
        "decision": "inconclusive",
        "difference": -0.2194270794105151,
        "estimand": "mean net contribution per assigned account, contact minus control",
        "interval_method": "normal approximation, unequal arm variances",
        "lower_95": -6.882298445609065,
        "minimum_detectable_effect_80_power": 9.518387665997928,
        "simulator_designed_effect": 2.70675,
        "standard_error": 3.3994241664278317,
        "upper_95": 6.443444286788035
      },
      "interpretation": "One simulated randomized experiment analysed by assigned arm, including failed deliveries. An interval spanning zero is inconclusive, not evidence of no effect. No real business impact is established.",
      "sample_ratio_check": {
        "alarm_threshold": 0.001,
        "chi_square": 3.38,
        "expected": {
          "0": 400.0,
          "1": 400.0
        },
        "observed": {
          "0": 374,
          "1": 426
        },
        "p_value": 0.0659921
      }
    },
    "retention": {
      "budget": 8,
      "contact_cost": 4,
      "profiles": [
        {
          "churn_risk": 0.65,
          "customer_id": "example-1",
          "eligible": true,
          "incremental_value": -2.0,
          "segment": "high_risk"
        },
        {
          "churn_risk": 0.65,
          "customer_id": "example-2",
          "eligible": true,
          "incremental_value": -2.0,
          "segment": "high_risk"
        },
        {
          "churn_risk": 0.35,
          "customer_id": "example-3",
          "eligible": true,
          "incremental_value": 14.0,
          "segment": "saveable"
        },
        {
          "churn_risk": 0.35,
          "customer_id": "example-4",
          "eligible": true,
          "incremental_value": 14.0,
          "segment": "saveable"
        },
        {
          "churn_risk": 0.1,
          "customer_id": "example-5",
          "eligible": true,
          "incremental_value": -3.0,
          "segment": "stable"
        },
        {
          "churn_risk": 0.1,
          "customer_id": "example-6",
          "eligible": true,
          "incremental_value": -3.0,
          "segment": "stable"
        }
      ],
      "risk_contacts": [
        "example-1",
        "example-2"
      ],
      "risk_expected_increment": -4.0,
      "source": "known simulator effects; oracle illustration, not learned causal estimates",
      "value_contacts": [
        "example-3",
        "example-4"
      ],
      "value_expected_increment": 28.0
    }
  },
  "candidate": {
    "brier": 0.16902006107806244,
    "confusion": {
      "fn": 2,
      "fp": 47,
      "tn": 27,
      "tp": 74
    },
    "cost_per_customer": 0.76,
    "log_loss": 0.4959357818758788,
    "n": 150,
    "prevalence": 0.5066666666666667,
    "roc_auc": 0.8372155049786629,
    "threshold": 0.15
  },
  "cohort_sizes": {
    "calibration": 150,
    "test": 150,
    "train": 300
  },
  "data_kind": "synthetic; no measured real-world impact",
  "execution": {
    "decisions": [
      {
        "cost": 4,
        "customer_id": "example-3",
        "decision_id": "contact-001",
        "status": "unknown"
      },
      {
        "cost": 4,
        "customer_id": "example-4",
        "decision_id": "contact-002",
        "status": "confirmed_not_delivered"
      }
    ],
    "lesson": "Unknown delivery retains its reservation; a repeated request does not reserve twice.",
    "remaining_budget": 4
  },
  "experiment_decision": "inconclusive",
  "experiment_gate": "Sample-ratio chi-square p >= 0.001 before reading lift; claim a simulated contact effect only if the 95% intention-to-treat interval excludes zero, otherwise inconclusive. No real campaign approved.",
  "gate": "Upper paired 95% bootstrap bound on candidate-minus-baseline log loss < -0.01, and candidate Brier <= baseline Brier. No real deployment approved.",
  "judge": {
    "confusion": {
      "fn": 4,
      "fp": 3,
      "tn": 17,
      "tp": 26
    },
    "false_acceptance_rate": 0.15,
    "limitation": "Fixed fixture arithmetic, not a measured judge or evidence of general reliability",
    "positive_means": "acceptable answer",
    "precision": 0.896551724137931,
    "recall": 0.8666666666666667,
    "source": "hand-constructed 50-case teaching fixture"
  },
  "lab_version": "1.1.0",
  "limitations": [
    "Entirely synthetic: predictive and causal findings do not establish real-world performance or business impact.",
    "Fixed independent-customer cohorts; real repeated users require grouped/time-aware resampling.",
    "One final test and predeclared gate. Changing the candidate after seeing this test requires fresh confirmatory data.",
    "Sigmoid calibration on a separate cohort is not a guarantee after population shift.",
    "Oracle retention effects and assumed price response are teaching assumptions, not learned causal estimates.",
    "The randomized experiment is far smaller than its power check requires; its interval, not its point estimate, is the result."
  ],
  "paired_log_loss_difference": {
    "draws": 1000,
    "lower_95": -0.289922080655043,
    "mean": -0.19981560856006947,
    "resampling_unit": "unique synthetic customer; paired predictions",
    "upper_95": -0.11539338673105394
  },
  "prototype_decision": "pass_simulated_gate",
  "seed": 42
}