{
  "format_version": "1.0",
  "dataset_title": "Aithra: Reciprocal Verification for Recursive Self-Improvement",
  "total_rows": 981,
  "split_note": "train is a dataset-loader container label; no new experimental train/test division has been introduced. Rows are experiment records, not training examples.",
  "units_note": "Query-work modeling, declared amortized setup charges, measured compilation seconds, and counted-work proxies have distinct units. None is an established neural-model speedup.",
  "configs": [
    {
      "config_name": "role_trials",
      "path": "data/role_trials.csv",
      "row_count": 480,
      "split": "train",
      "fields": [
        {
          "name": "seed",
          "dtype": "int64",
          "description": "Generator seed; reused across workload families and cost profiles."
        },
        {
          "name": "family",
          "dtype": "string",
          "description": "Synthetic workload family: aligned, disjoint, paired, rotated, drift, or mostly_valid."
        },
        {
          "name": "cost_kind",
          "dtype": "string",
          "description": "Common test cost profile: unit or heterogeneous."
        },
        {
          "name": "roles",
          "dtype": "int64",
          "description": "Number of verification roles."
        },
        {
          "name": "tests",
          "dtype": "int64",
          "description": "Number of tests/obligation positions."
        },
        {
          "name": "observations_per_role",
          "dtype": "int64",
          "description": "Source workload observations acquired for each role."
        },
        {
          "name": "source_acquisition_expected_work",
          "dtype": "float64",
          "description": "Expected query-work cost of all source observations using the initial order, reported separately."
        },
        {
          "name": "frozen_expected",
          "dtype": "float64",
          "description": "Target expected query work per call for the original test order, excluding setup."
        },
        {
          "name": "shared_expected",
          "dtype": "float64",
          "description": "Target expected query work per call for one learned shared order, excluding setup."
        },
        {
          "name": "separate_expected",
          "dtype": "float64",
          "description": "Target expected query work per call averaged over independently ordered roles, excluding setup."
        },
        {
          "name": "partition_expected",
          "dtype": "float64",
          "description": "Target expected query work per call for the learned partition of roles, excluding setup."
        },
        {
          "name": "partition_count",
          "dtype": "int64",
          "description": "Number of policies/groups selected by the partition compiler."
        },
        {
          "name": "shared_objective",
          "dtype": "float64",
          "description": "shared_expected + 0.35 (one declared amortized policy setup charge)."
        },
        {
          "name": "separate_objective",
          "dtype": "float64",
          "description": "separate_expected + roles * 0.35."
        },
        {
          "name": "partition_objective",
          "dtype": "float64",
          "description": "partition_expected + partition_count * 0.35."
        },
        {
          "name": "oracle_shared_expected",
          "dtype": "float64",
          "description": "Expected target query work for the shared order using the exact target distribution, excluding setup."
        },
        {
          "name": "oracle_separate_expected",
          "dtype": "float64",
          "description": "Expected target query work averaged over role-specific orders using exact target distributions, excluding setup."
        },
        {
          "name": "oracle_partition_objective",
          "dtype": "float64",
          "description": "Best partition objective using exact target distributions, including declared setup charges."
        },
        {
          "name": "compiler_work_proxy",
          "dtype": "int64",
          "description": "Compiler counted-work proxy, not CPU instruction count and not added to query-work lifecycle costs."
        },
        {
          "name": "compiler_seconds",
          "dtype": "float64",
          "description": "Recorded wall-clock seconds for partition compilation in the original environment."
        }
      ],
      "source": "research/results/role_trials.csv",
      "sha256": "0f5e9a306a1c9421a8424247532587c7e72f369b2388ec1b7d22c65a6b53eeeb"
    },
    {
      "config_name": "guarded_trials",
      "path": "data/guarded_trials.csv",
      "row_count": 480,
      "split": "train",
      "fields": [
        {
          "name": "seed",
          "dtype": "int64",
          "description": "Training generator seed; reused across workload families and cost profiles. Separate audit stream adds 982341."
        },
        {
          "name": "family",
          "dtype": "string",
          "description": "Synthetic workload family, matching role_trials conventions."
        },
        {
          "name": "cost_kind",
          "dtype": "string",
          "description": "Common test cost profile: unit or heterogeneous."
        },
        {
          "name": "audit_n",
          "dtype": "int64",
          "description": "Fresh audit call count, 4096 except zero when candidate and baseline orders are identical."
        },
        {
          "name": "individual_alpha",
          "dtype": "float64",
          "description": "Per-comparison statistical error level, 0.05/480; unrelated to reversal mixture alpha."
        },
        {
          "name": "lower_saving_bound",
          "dtype": "float64",
          "description": "Paired audit lower bound after extra setup, sampling radius, and actual audit work divided by deployment horizon; promote if > 1e-10."
        },
        {
          "name": "adopt",
          "dtype": "bool",
          "description": "Whether the candidate passed the fresh-audit promotion rule."
        },
        {
          "name": "shared_objective",
          "dtype": "float64",
          "description": "True target expected baseline query work per deployment call plus 0.35 setup charge."
        },
        {
          "name": "unguarded_objective",
          "dtype": "float64",
          "description": "True target expected candidate query work per deployment call plus all declared policy setup charges."
        },
        {
          "name": "guarded_objective",
          "dtype": "float64",
          "description": "Candidate objective if adopted, otherwise baseline objective."
        },
        {
          "name": "true_net_saving_if_adopted",
          "dtype": "float64",
          "description": "shared_objective - unguarded_objective; excludes source acquisition and audit work."
        },
        {
          "name": "audit_work",
          "dtype": "float64",
          "description": "Realized fresh-audit query work evaluating both baseline and candidate on paired sampled outcomes."
        },
        {
          "name": "training_acquisition_work",
          "dtype": "float64",
          "description": "Expected query-work cost of source data acquisition using original full policy."
        },
        {
          "name": "deployment_horizon",
          "dtype": "int64",
          "description": "Assumed future deployment calls, 100000."
        },
        {
          "name": "guarded_lifecycle_work",
          "dtype": "float64",
          "description": "training_acquisition_work + audit_work + deployment_horizon * guarded_objective."
        },
        {
          "name": "shared_lifecycle_work",
          "dtype": "float64",
          "description": "training_acquisition_work + deployment_horizon * shared_objective."
        },
        {
          "name": "unsafe_promotion",
          "dtype": "bool",
          "description": "adopt and unguarded_objective > shared_objective + 1e-9; evaluates deployment objective, not full lifecycle work."
        }
      ],
      "source": "research/results/guarded_trials.csv",
      "sha256": "d420dc6731a3f9bdd9b14dd6c18e73c63f28da57619f9f8a76a658b49b3b8a37"
    },
    {
      "config_name": "reversal",
      "path": "data/reversal.csv",
      "row_count": 21,
      "split": "train",
      "fields": [
        {
          "name": "alpha",
          "dtype": "float64",
          "description": "Mixture weight in q_alpha=(1-alpha)*p+alpha*q, ranging 0 to 1; not statistical alpha."
        },
        {
          "name": "true_gain",
          "dtype": "float64",
          "description": "Exact modeled expected query-work reduction of new order versus old order at q_alpha."
        },
        {
          "name": "robust_lower",
          "dtype": "float64",
          "description": "Worst-case saving under total-variation radius 0.05 computed by robust_promotion."
        },
        {
          "name": "promote",
          "dtype": "bool",
          "description": "Whether the robust worst-case bound authorizes the new policy."
        }
      ],
      "source": "research/results/reversal.csv",
      "sha256": "cd34c3fe78c218bfc70c1ea40c7bdea3d4bfe76e593a1e9c6ef42619b92c4606"
    }
  ],
  "dtype_note": "Field dtypes describe logical data. Depending on datasets/PyArrow versions, textual fields may be represented as string or large_string."
}