{
  "acceptance_version": "0.1",
  "object_path": "examples/research-data-business-data-sharing-through-data-marketplace",
  "manifest": "agent/tools.json",
  "published_results_path": "results.json",
  "notes": "Every block the published run reports is reproduced whole at the tool defaults, cross-linked to its recipe. The off-reference calls ask what the two robustness checks did not: whether the year association holds inside either era of the corpus. From 2010 onward (425 of 505 records) H1 and H2 are both null, and within the 80 pre-2010 records they are null too, though that tail holds only 11 Includes and can rule out only a large effect. The published association is a step in the inclusion rate (11 of 80 before 2010, 269 of 425 from 2010), not a gradient detectable inside either era. The 2010 cut is post hoc, taken at the published R2 bin boundary; any start year from 2009 to 2014 leaves both tests null, while starting at 2006 restores a significant H2 trend. H3's non-uniformity survives dropping its dominant group. abs_tolerance is 0 on every call: most values here are p-values at or below 1e-7, where the default 1e-9 absolute floor would loosen the 1e-6 relative contract to percent level or void it.",
  "calls": [
    {
      "call_id": "source-md5-published",
      "tool": "corpus_summary",
      "args": {},
      "select": "/source_md5",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/_meta/source_md5"
      },
      "recipe_id": "data-marketplace-taxonomy-source-md5",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The md5 here is digested from the fetched bytes, not echoed from the constant."
    },
    {
      "call_id": "n-records-published",
      "tool": "corpus_summary",
      "args": {},
      "select": "/n_records",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/_meta/n_records"
      },
      "recipe_id": "data-marketplace-taxonomy-records",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "n-include-published",
      "tool": "corpus_summary",
      "args": {},
      "select": "/n_include",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/_meta/n_include"
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "n-exclude-published",
      "tool": "corpus_summary",
      "args": {},
      "select": "/n_exclude",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/_meta/n_exclude"
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "normality-published",
      "tool": "normality_check",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/normality_shapiro"
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "h1-published",
      "tool": "year_by_decision_test",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/H1_mannwhitney_year_by_decision"
      },
      "recipe_id": "data-marketplace-taxonomy-h1-year",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "h1-scopus-only-published",
      "tool": "year_by_decision_test",
      "args": {
        "scopus_only": true
      },
      "select": "/p",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/R1_robustness_scopus_only/mannwhitney/p"
      },
      "recipe_id": "data-marketplace-taxonomy-r1-scopus",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The scopus_only parameter reproduces the R1 robustness value through the H1 tool."
    },
    {
      "call_id": "h1-from-2010-off-reference",
      "tool": "year_by_decision_test",
      "args": {
        "year_min": 2010
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Mann-Whitney U (two-sided)",
          "statistic_U": 21429.0,
          "p": 0.7111355491966639,
          "effect_size_rank_biserial": 0.021303974835573447,
          "auc_include_gt_exclude": 0.5106519874177867,
          "median_include": 2018.0,
          "median_exclude": 2018.0,
          "iqr_include": [
            2016.0,
            2019.0
          ],
          "iqr_exclude": [
            2015.0,
            2019.0
          ],
          "n_include": 269,
          "n_exclude": 156,
          "p_holm": 0.7111355491966639,
          "significant_holm": false
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The most consequential call in this file. From 2010 onward -- 425 of the 505 records, 84% of the corpus -- included and excluded records show no detectable difference in year: U = 21429, p = 0.711, rank-biserial r = 0.021, both medians 2018. The published H1 effect (r = 0.28, p = 5.3e-08) disappears once the 80 pre-2010 records, 11 of them included, are set aside. The cut is post hoc but not knife-edge: every start year from 2009 to 2014 gives p >= 0.50, and only a start of 2005 or earlier brings p below 0.05 (0.018 from 2005). Claim 1 is true of the corpus as screened, but what it measures is the contrast between the sparse early records and the 2010-2020 block, not a year difference within that decade."
    },
    {
      "call_id": "h1-before-2010-off-reference",
      "tool": "year_by_decision_test",
      "args": {
        "year_max": 2009
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Mann-Whitney U (two-sided)",
          "statistic_U": 315.5,
          "p": 0.37290437043462943,
          "effect_size_rank_biserial": -0.16864295125164686,
          "auc_include_gt_exclude": 0.4156785243741766,
          "median_include": 1998.0,
          "median_exclude": 2003.0,
          "iqr_include": [
            1995.5,
            2003.5
          ],
          "iqr_exclude": [
            1997.0,
            2005.0
          ],
          "n_include": 11,
          "n_exclude": 69,
          "p_holm": 0.37290437043462943,
          "significant_holm": false
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Inside the 80 pre-2010 records there is no detectable association (p = 0.373), and the point estimate runs the other way (r = -0.169; Include median 1998 against Exclude median 2003). With 11 Includes against 69 Excludes the test has little power, so this shows that the early tail does not carry the published direction on its own, not that the association there is zero. Neither era shows a detectable effect alone; pooling them does."
    },
    {
      "call_id": "h1-directional-off-reference",
      "tool": "year_by_decision_test",
      "args": {
        "alternative": "greater"
      },
      "select": "/p",
      "expect": {
        "kind": "literal",
        "value": 2.6380109747949184e-08
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The directional alternative (Include years stochastically greater) gives exactly half the two-sided p, as the normal approximation must when U lies on the hypothesised side of its mean. abs_tolerance is 0 so this p is held to the 1e-6 relative contract: under the default 1e-9 floor, dropping the continuity correction (p = 2.6334e-08, a difference of 4.6e-11) would still have matched."
    },
    {
      "call_id": "h2-published",
      "tool": "inclusion_trend_test",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/H2_cochran_armitage_trend"
      },
      "recipe_id": "data-marketplace-taxonomy-h2-trend",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "h2-scopus-only-published",
      "tool": "inclusion_trend_test",
      "args": {
        "scopus_only": true
      },
      "select": "/statistic_Z",
      "expect": {
        "kind": "published_pointer",
        "result_path": "/R1_robustness_scopus_only/cochran_armitage/statistic_Z"
      },
      "recipe_id": "data-marketplace-taxonomy-r1-scopus",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The scopus_only parameter reproduces the R1 trend statistic through the H2 tool."
    },
    {
      "call_id": "h2-from-2010-off-reference",
      "tool": "inclusion_trend_test",
      "args": {
        "year_min": 2010
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Cochran-Armitage trend (two-sided)",
          "statistic_Z": 0.4854797826063844,
          "p": 0.6273360505228642,
          "spearman_rho_year_include": 0.01800502455040276,
          "spearman_p": 0.7112931555353512,
          "direction": "inclusion proportion increases with year",
          "n": 425,
          "n_years": 11,
          "p_holm": 0.6273360505228642,
          "significant_holm": false
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Over 2010-2020 the inclusion rate shows no trend: Z = 0.49, p = 0.627, Spearman rho = 0.018, across 11 years and 425 records. The rise Claim 2 reports is real across 1980-2020, but its shape is a step -- 11 of 80 records included before 2010 (13.8%), 269 of 425 from 2010 (63.3%) -- not a gradient running through the period. Any start year from 2009 to 2014 leaves the trend null (p >= 0.31; the 2009 start gives p = 0.317); h2-from-2006-off-reference shows where it returns. Neither published robustness check can see this: Scopus-only and 5-year binning both keep the pre-2010 records whose contrast with the later block produces the result."
    },
    {
      "call_id": "h2-from-2006-off-reference",
      "tool": "inclusion_trend_test",
      "args": {
        "year_min": 2006
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Cochran-Armitage trend (two-sided)",
          "statistic_Z": 2.346880331670202,
          "p": 0.018931333954424916,
          "spearman_rho_year_include": 0.06682542464862605,
          "spearman_p": 0.16267928216779465,
          "direction": "inclusion proportion increases with year",
          "n": 438,
          "n_years": 15,
          "p_holm": 0.018931333954424916,
          "significant_holm": true
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Where the recent-era null ends. Starting at 2006 adds 13 records from 2006-2009, 2 of them included, and the trend is significant again (Z = 2.35, p = 0.019), although the Spearman correlation over the same 438 records is not (rho = 0.067, p = 0.163), nor is H1 from 2006 (p = 0.163). A significant trend inside a window needs some of the low-inclusion years before 2010; the 2010-2020 plateau alone has none."
    },
    {
      "call_id": "h2-before-2010-off-reference",
      "tool": "inclusion_trend_test",
      "args": {
        "year_max": 2009
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Cochran-Armitage trend (two-sided)",
          "statistic_Z": -0.5035225776421277,
          "p": 0.6145969107063792,
          "spearman_rho_year_include": -0.10103990731696091,
          "spearman_p": 0.372505405885578,
          "direction": "decreases",
          "n": 80,
          "n_years": 22,
          "p_holm": 0.6145969107063792,
          "significant_holm": false
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Within the 80 pre-2010 records the trend is null and its sign reverses (Z = -0.50, p = 0.615). With 11 Includes spread over 22 years the test has little power, so this shows only that the early tail does not carry the published direction on its own."
    },
    {
      "call_id": "h2-rank-scores-off-reference",
      "tool": "inclusion_trend_test",
      "args": {
        "scores": "rank"
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Cochran-Armitage trend (two-sided, rank scores)",
          "statistic_Z": 7.2070542711386025,
          "p": 5.717522045145591e-13,
          "spearman_rho_year_include": 0.2424084946176082,
          "spearman_p": 3.459328076095638e-08,
          "direction": "inclusion proportion increases with year",
          "n": 505,
          "n_years": 33,
          "p_holm": 1.1435044090291183e-12,
          "significant_holm": true
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Scoring years 1..33 instead of by calendar value removes the extra leverage the isolated early years (1980 and 1984, one record each) get from their distance to the rest of the series. Z moves from 7.18 to 7.21: the full-corpus result does not depend on the score choice."
    },
    {
      "call_id": "h3-published",
      "tool": "exclusion_reason_test",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/H3_chisquare_exclusion_groups"
      },
      "recipe_id": "data-marketplace-taxonomy-h3-exclusion",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "h3-without-off-topic-off-reference",
      "tool": "exclusion_reason_test",
      "args": {
        "groups": [
          2,
          3,
          4,
          5
        ]
      },
      "expect": {
        "kind": "literal",
        "value": {
          "test": "Chi-square goodness-of-fit (equiprobable)",
          "statistic_chi2": 31.184615384615388,
          "p": 7.772634320021047e-07,
          "dof": 3,
          "effect_size_cohens_w": 0.6926494882763341,
          "observed": {
            "2": 33,
            "3": 20,
            "4": 5,
            "5": 7
          },
          "expected_each": 16.25,
          "labels": {
            "2": "Marketplace peripheral; emphasis elsewhere",
            "3": "Workshop/proceeding description, not a research paper",
            "4": "No abstract available",
            "5": "Not in English"
          },
          "n": 65,
          "p_holm": 7.772634320021047e-07,
          "significant_holm": true
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "An equiprobable null over five reasons, one of which is 'off-topic', is close to a straw man for a broad keyword query. Setting group 1 aside, the four remaining reasons are still far from uniform (chi2(3) = 31.2, p = 7.8e-07, w = 0.69; 33 of 65 are 'marketplace peripheral'). Claim 3's non-uniformity is therefore not only the off-topic group's dominance, though an equiprobable null stays a weak benchmark for administrative reasons such as 'no abstract' and 'not in English'."
    },
    {
      "call_id": "holm-published",
      "tool": "holm_correction",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/holm_correction"
      },
      "recipe_id": "data-marketplace-taxonomy-holm",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "holm-all-six-tests-off-reference",
      "tool": "holm_correction",
      "args": {
        "p_values": [
          5.276021949589837e-08,
          6.89933974521531e-13,
          1.1552958846495774e-80,
          1.7041808247101982e-07,
          2.7639903792579752e-12,
          3.320309561764601e-13
        ]
      },
      "expect": {
        "kind": "literal",
        "value": {
          "method": "Holm-Bonferroni",
          "family": [
            "p_values[0]",
            "p_values[1]",
            "p_values[2]",
            "p_values[3]",
            "p_values[4]",
            "p_values[5]"
          ],
          "raw_p": [
            5.276021949589837e-08,
            6.89933974521531e-13,
            1.1552958846495774e-80,
            1.7041808247101982e-07,
            2.7639903792579752e-12,
            3.320309561764601e-13
          ],
          "corrected_p": [
            1.0552043899179674e-07,
            2.759735898086124e-12,
            6.931775307897464e-80,
            1.7041808247101982e-07,
            8.291971137773925e-12,
            1.6601547808823005e-12
          ],
          "reject_at_0.05": [
            true,
            true,
            true,
            true,
            true,
            true
          ]
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Holm over the six claim-bearing tests -- the three primaries, R1's two re-runs and R2 -- rather than the declared family of three; the Shapiro-Wilk checks and the Spearman p reported beside H2 are left out. Every member still rejects; the largest corrected p is 1.7e-07. The inputs are literals and the arithmetic is exact. abs_tolerance is 0, as on every call here: four of the six values in each list sit below the default 1e-9 absolute floor, which would make their comparison vacuous, and the other two would be checked only to 0.6-1.9% relative."
    },
    {
      "call_id": "r1-published",
      "tool": "subset_robustness",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/R1_robustness_scopus_only"
      },
      "recipe_id": "data-marketplace-taxonomy-r1-scopus",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "r1-scopus-from-2010-off-reference",
      "tool": "subset_robustness",
      "args": {
        "scopus_only": true,
        "year_min": 2010
      },
      "expect": {
        "kind": "literal",
        "value": {
          "description": "Primary tests re-run on Scopus-only records (9 hand-seeded core Includes dropped), publication years 2010 onward",
          "n": 416,
          "n_include": 260,
          "n_exclude": 156,
          "mannwhitney": {
            "statistic_U": 20483.0,
            "p": 0.8629252929730329,
            "effect_size_rank_biserial": 0.010009861932938957,
            "median_include": 2018.0,
            "median_exclude": 2018.0
          },
          "cochran_armitage": {
            "statistic_Z": 0.3116490936855504,
            "p": 0.7553072196572339
          }
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Both robustness worries at once: drop the 9 hand-seeded core articles (all dated 2014-2020, so all inside this window) and the pre-2010 records. Both year tests are null (Mann-Whitney p = 0.863, Cochran-Armitage p = 0.755). The published R1 check addressed retrieval provenance, not era, so it kept the pre-2010 records that produce the association and could not have shown this."
    },
    {
      "call_id": "r1-scopus-before-2010-off-reference",
      "tool": "subset_robustness",
      "args": {
        "scopus_only": true,
        "year_max": 2009
      },
      "expect": {
        "kind": "literal",
        "value": {
          "description": "Primary tests re-run on Scopus-only records (no non-Scopus records in this window to drop), publication years up to 2009",
          "n": 80,
          "n_include": 11,
          "n_exclude": 69,
          "mannwhitney": {
            "statistic_U": 315.5,
            "p": 0.37290437043462943,
            "effect_size_rank_biserial": -0.16864295125164686,
            "median_include": 1998.0,
            "median_exclude": 2003.0
          },
          "cochran_armitage": {
            "statistic_Z": -0.5035225776421277,
            "p": 0.6145969107063792
          }
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Scopus-only up to 2009 is the same 80 records as the pre-2010 window alone, because all 9 core articles date from 2014-2020; the numbers equal h1-before-2010 and h2-before-2010. The description counts the non-Scopus records inside the window, so it says none were dropped rather than quoting the corpus-wide 9."
    },
    {
      "call_id": "r2-published",
      "tool": "binned_trend_test",
      "args": {},
      "expect": {
        "kind": "published_pointer",
        "result_path": "/R2_sensitivity_5yr_bins"
      },
      "recipe_id": "data-marketplace-taxonomy-r2-bins",
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true
    },
    {
      "call_id": "r2-shifted-equal-bins-off-reference",
      "tool": "binned_trend_test",
      "args": {
        "bin_edges": [
          1977,
          1982,
          1987,
          1992,
          1997,
          2002,
          2007,
          2012,
          2017,
          2022
        ]
      },
      "expect": {
        "kind": "literal",
        "value": {
          "description": "Trend test on publication-year bins with edges [1977, 1982, 1987, 1992, 1997, 2002, 2007, 2012, 2017, 2022]",
          "bins": [
            "(1977, 1982]",
            "(1982, 1987]",
            "(1987, 1992]",
            "(1992, 1997]",
            "(1997, 2002]",
            "(2002, 2007]",
            "(2007, 2012]",
            "(2012, 2017]",
            "(2017, 2022]"
          ],
          "n_per_bin": [
            1,
            1,
            6,
            15,
            16,
            31,
            33,
            178,
            224
          ],
          "include_per_bin": [
            0,
            0,
            0,
            5,
            2,
            2,
            17,
            111,
            143
          ],
          "inclusion_rate_per_bin": [
            0.0,
            0.0,
            0.0,
            0.3333333333333333,
            0.125,
            0.06451612903225806,
            0.5151515151515151,
            0.6235955056179775,
            0.6383928571428571
          ],
          "cochran_armitage": {
            "statistic_Z": 6.887294172173657,
            "p": 5.6863558816254445e-12
          }
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "Five-year bins shifted two years earlier and of truly equal width (the published last bin, (2014, 2020], is six years wide). Z = 6.89 against the published 7.28: the binned trend does not depend on where the bins are anchored. The per-bin rates show the shape directly -- 0.52, 0.62, 0.64 across the last three bins against 0.00-0.33 in the six before -- a plateau reached around 2010, not a steady climb."
    },
    {
      "call_id": "r2-scopus-only-off-reference",
      "tool": "binned_trend_test",
      "args": {
        "scopus_only": true
      },
      "select": "/cochran_armitage",
      "expect": {
        "kind": "literal",
        "value": {
          "statistic_Z": 7.13207663081361,
          "p": 9.88658017454045e-13
        }
      },
      "rel_tolerance": 1e-06,
      "abs_tolerance": 0.0,
      "enabled": true,
      "notes": "The published bins on Scopus-only records: Z = 7.13, so R1 and R2 compose without changing the conclusion."
    }
  ]
}