{
  "arithmetic": "Exact integer counts and rational rates; finite decimal output is convenience only.",
  "assertionTested": "The reader retains matched task clusters and every episode, reconstructs two failure-rate estimands, and calculates the complete empirical paired task bootstrap distribution for these retained counts.",
  "caps": {
    "episodesBothArms": 65536,
    "exhaustiveDraws": 65536,
    "inputBytes": 2097152,
    "reportBytes": 134217728,
    "sampledDraws": [
      32,
      4096
    ],
    "tasks": 64
  },
  "contrast": "attack minus benign, recomputed jointly within every paired draw",
  "estimands": {
    "episode_weighted_resolved": "ratio of summed failure counts to summed failure plus success counts, separately in each arm",
    "task_weighted_resolved": "equal-weight mean of all declared per-task failure proportions, separately in each arm"
  },
  "evidenceKind": "synthetic-illustration",
  "format": "probity-task-grouped-rates-contract-v1",
  "input": "input.json",
  "labels": [
    "failure",
    "success",
    "unresolved"
  ],
  "missingData": {
    "emptyArm": "retain declared task and return null rate; never substitute zero",
    "undefinedDraw": "retain null statistic; do not discard draw; block affected percentile interval",
    "unresolved": "retain count; exclude only from resolved denominators; also report all-attempt worst-case bounds"
  },
  "nonClaims": [
    "No empirical robot or model result.",
    "No new Lab registry record or producer acceptance.",
    "No verified independent operation or custody.",
    "No exact population confidence coverage or causal attack effect."
  ],
  "output": "report.json",
  "reader": "../../tools/check_grouped_rates.py",
  "requiredFields": {
    "counts": [
      "total",
      "failure",
      "success",
      "unresolved"
    ],
    "episode": [
      "episodeId",
      "outcome"
    ],
    "matchedArm": [
      "conditions",
      "intervention"
    ],
    "matching": [
      "attack",
      "benign",
      "outcomeCriteria"
    ],
    "monteCarloAdditional": [
      "seed",
      "replicates"
    ],
    "population": [
      "tasks",
      "episodes",
      "selection",
      "taskIndependence",
      "independenceBasis",
      "residualDependence"
    ],
    "resampling": [
      "mode",
      "confidenceLevel"
    ],
    "root": [
      "schemaVersion",
      "evidenceKind",
      "population",
      "matching",
      "resampling",
      "tasks"
    ],
    "task": [
      "taskId",
      "attack",
      "benign"
    ],
    "taskArm": [
      "episodes",
      "counts"
    ]
  },
  "samplingUnit": "paired-task-cluster",
  "uncertainty": {
    "assumptions": [
      "Independent exchangeable task clusters for generalization; both arms and all episodes are retained together.",
      "The declared tasks and episode population match the target estimand and the rubric is appropriate.",
      "A resolved-only rate can be biased by nonrandom unresolved labels."
    ],
    "drawMode": "exhaustive",
    "draws": 256,
    "interpretation": "Exact empirical resampling summary. Nominal confidence coverage is not established; four toy tasks do not support an empirical inference.",
    "method": "paired-task-bootstrap-percentile-v1",
    "nominalLevel": "0.95",
    "quantileRule": "left inverse empirical CDF, rank max(1, ceil(p * B)), no interpolation"
  }
}