{
  "schema_version": "1.0",
  "generated_utc": "2026-10-27T18:06:24.511077+00:00",
  "scope": {
    "catalogue_tasks": 31,
    "evaluated_tasks": 24,
    "deferred_tasks": 7,
    "domains": 10,
    "open_models": 5,
    "full_sample_task_model_cells": 120,
    "cpu_reference_validation_tasks": 24,
    "cpu_reference_validation_modes": 27,
    "api_models": 4
  },
  "accepted_input_snapshots": [
    "final_numeric_long_24_closeout_20261028.csv",
    "final_metric_status_24_closeout_20261028.csv",
    "FINAL_FIXED_SCOPE_COMPUTE_CLOSED_20261028.json",
    "final_baseline_paper_artifacts_20261028/full_condition_scores.csv",
    "task_catalog.json",
    "selected_tasks_24.json",
    "cpu_validation_24_status.json",
    "table2_values_20261030.json (main results table, paper commit 5a0f491)",
    "runs/probes_20261028/analysis/paper_body_revision_20261029/TABLE2_COMPLETE_UNIFORM_INTEGRATION_20261029.json",
    "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_REFERENCE_ROW_ALIGNMENT_20261029.json",
    "runs/probes_20261028/analysis/paper_revision_b_20261029/EXISTING_BANDIT_REFERENCE_TARGET_ALIGNMENT_20261029.json"
  ],
  "counts": {
    "private_numeric_rows": 9178,
    "public_rows_from_private_numeric_table": 1830,
    "gaussian_ts_condition_metric_rows_added": 16,
    "public_result_rows": 1961,
    "excluded_runtime_strata_rows": 7342,
    "pending_protocol_rows_not_displayed_as_complete": 6,
    "bandit_condition_download_rows": 88,
    "metric_status_rows": 1080,
    "public_value_status_counts": {
      "finite": 1883,
      "defined_positive_infinity": 16,
      "undefined": 62
    },
    "public_models": {
      "laya": 365,
      "intern_decision_4b": 394,
      "llama8b": 356,
      "minitaur8b": 381,
      "qwen35_4b": 339,
      "gaussian_ts": 16,
      "jev_1_14": 29,
      "claude_opus_5_5": 27,
      "claude_sonnet_5_5": 27,
      "claude_haiku_5_5": 27
    },
    "table2_marker_rows": {
      "n.a.": 23,
      "canc.": 2,
      "part.": 4,
      "n.e.": 26,
      "n.f.": 6
    },
    "table2_cells": 297
  },
  "export_rules": [
    "Include every accepted task/experiment/condition aggregate in the numeric table regardless of outcome, sign, Infinity or undefined value.",
    "Exclude fine-grained per-execution-file/input-session runtime strata from the public presentation bundle. Original private results are retained unchanged.",
    "Pending protocol-resolution values are not labelled completed results; all pending availability fields remain in metric_status.json.",
    "Add the accepted eight Gaussian Thompson conditions with reward and pseudo-regret. Preserve the separate disclosed classical-policy protocol.",
    "Do not pool reward distributions, arm counts, horizons, native/sample interfaces, human prediction, event probability or autonomous return.",
    "Retain original decimal text in value_raw and CSV; undefined is null, positive infinity is the string Infinity.",
    "Remove original absolute runtime paths, host identifiers, GPU UUIDs, private logs and credential-bearing fields. Source-row identifiers and JSON pointers preserve mapping to the accepted private inventory.",
    "No participant/trial responses, original transcripts, per-participant identifiers, gated data or model weights are included."
  ],
  "data_and_code_conditions": {
    "code_license": "Original S1CogBench project software is MIT licensed. Third-party code, models and source data retain their original terms.",
    "upstream_data": "Original provider terms apply independently. No blanket redistribution permission is asserted.",
    "psych101_test": "Provider access/conditions required; not redistributed.",
    "cpu_example": "400 newly generated urn queries using a pinned public CogBench source; analytical Bayes, no GPU or model weights.",
    "remaining_acquisition": "python scripts/fetch_public_sources.py all fetches every public source at its pinned revision into the layout the loaders read. Psych-101-test stays gated at its provider, and Julia 1.4.1 is needed for Mouselab autonomous.",
    "code_license_spdx": "MIT"
  },
  "files": [
    "tasks.json",
    "results.json",
    "results.csv",
    "bandit_condition_scores.csv",
    "metric_status.json",
    "result_schema.json",
    "grouping_guide.md",
    "api_time_cost_per_target.csv",
    "table2.csv",
    "table2.json",
    "minitaur_nonliteral_reparsed.csv",
    "rlwm_curves.csv",
    "cpu_validation_24_status.json"
  ],
  "limitations": [
    "120/120 means one full-sample accepted named protocol per selected task/model, not all modes or all metrics.",
    "35 needs_protocol_resolution metric fields and one all-invalid undefined field remain explicit.",
    "The 24-task/27-mode CPU execution check is not a rerun of five neural models or independent acquisition of gated data.",
    "Runtime measurements have different hosts/interfaces and cannot be pooled into an unconditional speed leaderboard.",
    "API rows keep the provider readout: native typed choice (Jev), generated FINAL line (Claude, human-NLL targets n.a.) or stated probabilities (Claude Haiku on four modes). They are not internal probabilities of open weights.",
    "Nine Uniform values (autonomous modes) are analytic expectations that no single run reproduces."
  ]
}