{
  "acceptance_version": "0.1",
  "object_path": "examples/dataset-belonging-to-the-help-desk-log-of-an-italian-company",
  "manifest": "agent/tools.json",
  "published_results_path": "results.json",
  "notes": "The published run is four calls to one tool. The off-reference calls reach the combinations it did not report, and the first of them materially qualifies claim 1.",
  "calls": [
    {
      "call_id": "meta-published",
      "tool": "dataset_summary",
      "args": {},
      "expect": {"kind": "published_pointer", "result_path": "/meta"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "primary-published",
      "tool": "tier_comparison",
      "args": {},
      "expect": {"kind": "published_pointer", "result_path": "/primary"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "r1-throughput-published",
      "tool": "tier_comparison",
      "args": {"outcome": "throughput_h"},
      "expect": {"kind": "published_pointer", "result_path": "/robustness/R1_throughput"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "r2-drop-ambiguous-published",
      "tool": "tier_comparison",
      "args": {"drop_ambiguous": true},
      "expect": {"kind": "published_pointer", "result_path": "/robustness/R2_drop_ambiguous"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "r3-first-observed-published",
      "tool": "tier_comparison",
      "args": {"grouping": "service_level_first"},
      "expect": {"kind": "published_pointer", "result_path": "/robustness/R3_first_observed"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "include-excluded-tier-off-reference",
      "tool": "tier_comparison",
      "args": {"tiers": ["Value 1", "Value 2", "Value 3", "Value 4"]},
      "expect": {
        "kind": "literal",
        "value": {
          "H": 7.123001732312447,
          "df": 3,
          "p": 0.06807892727557836,
          "epsilon_squared": 0.0009031767212075458,
          "n": 4569,
          "k": 4,
          "group_n": {"Value 1": 115, "Value 2": 3527, "Value 3": 924, "Value 4": 3},
          "group_median_h": {"Value 1": 354.70666666666665, "Value 2": 140.4713888888889, "Value 3": 160.51430555555555, "Value 4": 309.995}
        }
      },
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true,
      "notes": "The most consequential call in this file. Value 4 holds 3 cases of 4,569 and is excluded a priori. Putting it back moves the primary test from p = 0.029 to p = 0.068 -- from significant to not, on the conventional threshold. The exclusion was declared before the analysis, so this is not p-hacking; but the headline rests on three cases and on an epsilon-squared of 0.0011, and claim 1 should say so."
    },
    {
      "call_id": "drop-smallest-tier-off-reference",
      "tool": "tier_comparison",
      "args": {"tiers": ["Value 2", "Value 3"]},
      "select": "/p",
      "expect": {"kind": "literal", "value": 0.047263894310011115},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true,
      "notes": "Dropping Value 1 -- 115 cases with a median resolve time more than twice the other tiers' -- leaves the two large tiers differing at p = 0.047. The tier effect is not carried solely by the small slow group, but it clears 0.05 by a similarly thin margin."
    },
    {
      "call_id": "pairwise-published",
      "tool": "pairwise_comparisons",
      "args": {},
      "expect": {"kind": "published_pointer", "result_path": "/pairwise"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "pairwise-throughput-off-reference",
      "tool": "pairwise_comparisons",
      "args": {"outcome": "throughput_h"},
      "select": "/0/p_holm",
      "expect": {"kind": "literal", "value": 0.19581317640527854},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true,
      "notes": "The same pair on throughput rather than resolve time corrects to 0.196, against 0.142 on the published outcome. Neither survives Holm, so the Value 1 vs Value 2 contrast is not significant on either outcome once the family is corrected."
    },
    {
      "call_id": "effect-size-published",
      "tool": "effect_size",
      "args": {"h": 7.048836129897784, "n": 4566, "k": 3},
      "select": "/epsilon_squared",
      "expect": {"kind": "published_pointer", "result_path": "/primary/epsilon_squared"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "effect-size-scaled-off-reference",
      "tool": "effect_size",
      "args": {"h": 7.048836129897784, "n": 200, "k": 3},
      "select": "/epsilon_squared",
      "expect": {"kind": "literal", "value": 0.02562860979643545},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true,
      "notes": "The same H on 200 cases rather than 4,566 gives an epsilon-squared of 0.026 instead of 0.0011 -- a 23-fold difference. The vanishing effect size is a statement about the sample size, not about the tiers, which is exactly why the significant p deserves the caveat."
    },
    {
      "call_id": "holm-published",
      "tool": "holm",
      "args": {},
      "select": "/0",
      "expect": {"kind": "published_pointer", "result_path": "/pairwise/0/p_holm"},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true
    },
    {
      "call_id": "holm-ranked-family-off-reference",
      "tool": "holm",
      "args": {"p_values": [0.01, 0.04, 0.2]},
      "expect": {"kind": "literal", "value": [0.03, 0.08, 0.2]},
      "rel_tolerance": 1e-6,
      "abs_tolerance": 1e-9,
      "enabled": true,
      "notes": "A family where the step-down separates the members, which the published family's two tied corrections never show."
    }
  ]
}