{
 "tool": "surface_baseline",
 "version": "0.1",
 "tool_sha256": "8b27ce5a760612348ba29a47ff4a7fe3ca0efa31f21122ba0f772dbfc2802e2d",
 "nulltest_sha256": "0e103dc41a3c208d2652646f2b8acc8b701ed7799402ca1ea0cb856de5ddf46c",
 "nulltest_version": "1.4",
 "generated": "2026-10-25",
 "folds": 5,
 "seed": 20260724,
 "rows": [
  {
   "name": "aperture/verifydomain-claims-147",
   "label_kind": "true_vs_false_statement",
   "status": "scored",
   "sha256": "ed742e764a2f2e14f71ae6f4da5e7a88b2870529e75a6646ed804b709aa54d14",
   "items_total": 147,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 147,
   "n_label1": 36,
   "n_label0": 111,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.8844,
   "accuracy_ci95": [
    0.8213,
    0.9312
   ],
   "balanced_accuracy": 0.792,
   "auroc": 0.9469,
   "majority_accuracy": 0.7551,
   "stratum_prior_accuracy": 0.7551,
   "probe_auroc_held_out": {
    "lexicality": 0.967,
    "words": 0.965,
    "length": 0.64,
    "shape": 0.778,
    "alpha_first": 0.668,
    "alpha_last": 0.575
   },
   "headline": "surface baseline: 88.4% held-out accuracy [82.1%, 93.1%] from six probes, against 75.5% for the majority class; balanced accuracy 79.2%, AUROC 0.947."
  },
  {
   "name": "aperture/grounding-coordinate",
   "label_kind": "real_vs_fabricated_entity",
   "status": "scored",
   "sha256": "3e1dcce261a7f9dd7b4a09bcd0943f130b2a1ef7c3b71d7139e3f2f035dd9608",
   "items_total": 162,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 162,
   "n_label1": 81,
   "n_label0": 81,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 6,
   "accuracy": 0.4074,
   "accuracy_ci95": [
    0.331,
    0.4873
   ],
   "balanced_accuracy": 0.4074,
   "auroc": 0.3728,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.516,
    "words": 0.573,
    "length": 0.588,
    "shape": 0.536,
    "alpha_first": 0.534,
    "alpha_last": 0.537
   },
   "headline": "surface baseline: 40.7% held-out accuracy [33.1%, 48.7%] from six probes, against 50.0% for the majority class and 50.0% for knowing each stratum's label rate; balanced accuracy 40.7%, AUROC 0.373. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "aperture/photon-entities",
   "label_kind": "real_vs_fabricated_entity",
   "status": "scored",
   "sha256": "f97e87443535a036dcb2cd0240bbec6dec4025f8d618687d86f6cafc63a088cb",
   "items_total": 103,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 103,
   "n_label1": 27,
   "n_label0": 76,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 7,
   "accuracy": 0.8252,
   "accuracy_ci95": [
    0.7379,
    0.893
   ],
   "balanced_accuracy": 0.7144,
   "auroc": 0.8465,
   "majority_accuracy": 0.7379,
   "stratum_prior_accuracy": 0.7379,
   "probe_auroc_held_out": {
    "lexicality": 0.84,
    "words": 0.712,
    "length": 0.613,
    "shape": 0.632,
    "alpha_first": 0.707,
    "alpha_last": 0.52
   },
   "headline": "surface baseline: 82.5% held-out accuracy [73.8%, 89.3%] from six probes, against 73.8% for the majority class and 73.8% for knowing each stratum's label rate; balanced accuracy 71.4%, AUROC 0.847."
  },
  {
   "name": "aperture/prominence-people-v2",
   "label_kind": "known_vs_unknown_real_entity",
   "status": "scored",
   "sha256": "d93ad2b7975d10f005507ef9c33dbc7afc1b0ce240a5d956bbad7d025b01bd91",
   "items_total": 156,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 156,
   "n_label1": 78,
   "n_label0": 78,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.4872,
   "accuracy_ci95": [
    0.4065,
    0.5684
   ],
   "balanced_accuracy": 0.4872,
   "auroc": 0.4799,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.554,
    "words": 0.538,
    "length": 0.506,
    "shape": 0.544,
    "alpha_first": 0.521,
    "alpha_last": 0.5
   },
   "headline": "surface baseline: 48.7% held-out accuracy [40.6%, 56.8%] from six probes, against 50.0% for the majority class; balanced accuracy 48.7%, AUROC 0.480. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "aperture/prominence-people-v1",
   "label_kind": "known_vs_unknown_real_entity",
   "status": "scored",
   "sha256": "fd881034ccd02904dfae2296435ebc651ccce63b8d408333f018550ceb154322",
   "items_total": 168,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 168,
   "n_label1": 84,
   "n_label0": 84,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.619,
   "accuracy_ci95": [
    0.5411,
    0.6928
   ],
   "balanced_accuracy": 0.619,
   "auroc": 0.6321,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.546,
    "words": 0.536,
    "length": 0.528,
    "shape": 0.602,
    "alpha_first": 0.561,
    "alpha_last": 0.661
   },
   "headline": "surface baseline: 61.9% held-out accuracy [54.1%, 69.3%] from six probes, against 50.0% for the majority class; balanced accuracy 61.9%, AUROC 0.632."
  },
  {
   "name": "aperture/postcutoff-trials-v1",
   "label_kind": "known_vs_unknown_real_entity",
   "status": "scored",
   "sha256": "54a6dbd8569af454a81521f5916886e0abfb3c14e4f05bef9cb6ec2f9db236a3",
   "items_total": 702,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 702,
   "n_label1": 351,
   "n_label0": 351,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.4587,
   "accuracy_ci95": [
    0.4214,
    0.4964
   ],
   "balanced_accuracy": 0.4587,
   "auroc": 0.4643,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.526,
    "words": 0.532,
    "length": 0.5,
    "shape": 0.503,
    "alpha_first": 0.512,
    "alpha_last": 0.529
   },
   "headline": "surface baseline: 45.9% held-out accuracy [42.1%, 49.6%] from six probes, against 50.0% for the majority class; balanced accuracy 45.9%, AUROC 0.464. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "bielik-entity-familiarity-v2",
   "label_kind": "real_vs_fabricated_entity",
   "status": "scored",
   "sha256": "02546469b730cb173e16dabba926e08e6c5094a734cdb7efb383d2c1322d8573",
   "items_total": 1440,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 1440,
   "n_label1": 240,
   "n_label0": 1200,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 4,
   "accuracy": 0.9,
   "accuracy_ci95": [
    0.8833,
    0.915
   ],
   "balanced_accuracy": 0.735,
   "auroc": 0.9258,
   "majority_accuracy": 0.8333,
   "stratum_prior_accuracy": 0.8333,
   "probe_auroc_held_out": {
    "lexicality": 0.927,
    "words": 0.828,
    "length": 0.572,
    "shape": 0.576,
    "alpha_first": 0.576,
    "alpha_last": 0.57
   },
   "headline": "surface baseline: 90.0% held-out accuracy [88.3%, 91.5%] from six probes, against 83.3% for the majority class and 83.3% for knowing each stratum's label rate; balanced accuracy 73.5%, AUROC 0.926."
  },
  {
   "name": "hallucinated-references/gpt-4",
   "label_kind": "real_vs_fabricated_entity",
   "status": "scored",
   "sha256": "3fd32d5e9dc1657a381b94f88817b64b4a0e7bb007107f4db18b3206cdc7a22f",
   "items_total": 1000,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 1000,
   "n_label1": 468,
   "n_label0": 532,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.779,
   "accuracy_ci95": [
    0.752,
    0.8044
   ],
   "balanced_accuracy": 0.7778,
   "auroc": 0.8592,
   "majority_accuracy": 0.532,
   "stratum_prior_accuracy": 0.532,
   "probe_auroc_held_out": {
    "lexicality": 0.836,
    "words": 0.833,
    "length": 0.777,
    "shape": 0.684,
    "alpha_first": 0.534,
    "alpha_last": 0.541
   },
   "headline": "surface baseline: 77.9% held-out accuracy [75.2%, 80.4%] from six probes, against 53.2% for the majority class; balanced accuracy 77.8%, AUROC 0.859."
  },
  {
   "name": "geometry-of-truth/cities",
   "label_kind": "true_vs_false_statement",
   "status": "scored",
   "sha256": "f2560a0c1758c69935b50125111b574b71896bb75930dd87b1d8c973c1fa086e",
   "items_total": 1496,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 1496,
   "n_label1": 748,
   "n_label0": 748,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.8082,
   "accuracy_ci95": [
    0.7873,
    0.8278
   ],
   "balanced_accuracy": 0.8082,
   "auroc": 0.8712,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.765,
    "words": 0.554,
    "length": 0.529,
    "shape": 0.504,
    "alpha_first": 0.5,
    "alpha_last": 0.523
   },
   "headline": "surface baseline: 80.8% held-out accuracy [78.7%, 82.8%] from six probes, against 50.0% for the majority class; balanced accuracy 80.8%, AUROC 0.871."
  },
  {
   "name": "geometry-of-truth/sp_en_trans",
   "label_kind": "true_vs_false_statement",
   "status": "scored",
   "sha256": "9afa97c1b8af539a05f47655a80a5c7832c6fdb914885fd24cb4bcccbe2795b5",
   "items_total": 354,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 354,
   "n_label1": 177,
   "n_label0": 177,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.4972,
   "accuracy_ci95": [
    0.4439,
    0.5505
   ],
   "balanced_accuracy": 0.4972,
   "auroc": 0.4837,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.508,
    "words": 0.531,
    "length": 0.519,
    "shape": 0.517,
    "alpha_first": 0.5,
    "alpha_last": 0.542
   },
   "headline": "surface baseline: 49.7% held-out accuracy [44.4%, 55.0%] from six probes, against 50.0% for the majority class; balanced accuracy 49.7%, AUROC 0.484. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "truth-is-universal/facts",
   "label_kind": "true_vs_false_statement",
   "status": "scored",
   "sha256": "303ec926c2710a13a6166e204a24ea013bb2991b8761e2899739fddcb923aee1",
   "items_total": 561,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 561,
   "n_label1": 282,
   "n_label0": 279,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.59,
   "accuracy_ci95": [
    0.548,
    0.631
   ],
   "balanced_accuracy": 0.59,
   "auroc": 0.625,
   "majority_accuracy": 0.5027,
   "stratum_prior_accuracy": 0.5027,
   "probe_auroc_held_out": {
    "lexicality": 0.508,
    "words": 0.526,
    "length": 0.565,
    "shape": 0.566,
    "alpha_first": 0.503,
    "alpha_last": 0.539
   },
   "headline": "surface baseline: 59.0% held-out accuracy [54.8%, 63.1%] from six probes, against 50.3% for the majority class; balanced accuracy 59.0%, AUROC 0.625."
  },
  {
   "name": "truthfulqa/answers",
   "label_kind": "right_vs_wrong_answer",
   "status": "scored",
   "sha256": "b8d8ef1e12f98b4f2a9f47abc9765da0640b182b6c5d9b92f0c1a1f2f1e02e5c",
   "items_total": 6028,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 6028,
   "n_label1": 3251,
   "n_label0": 2777,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 37,
   "accuracy": 0.7208,
   "accuracy_ci95": [
    0.7093,
    0.7321
   ],
   "balanced_accuracy": 0.7122,
   "auroc": 0.796,
   "majority_accuracy": 0.5393,
   "stratum_prior_accuracy": 0.5513,
   "probe_auroc_held_out": {
    "lexicality": 0.792,
    "words": 0.787,
    "length": 0.565,
    "shape": 0.582,
    "alpha_first": 0.502,
    "alpha_last": 0.504
   },
   "headline": "surface baseline: 72.1% held-out accuracy [70.9%, 73.2%] from six probes, against 53.9% for the majority class and 55.1% for knowing each stratum's label rate; balanced accuracy 71.2%, AUROC 0.796."
  },
  {
   "name": "halueval/qa",
   "label_kind": "right_vs_wrong_answer",
   "status": "scored",
   "sha256": "89ed139ec5e3a3169a0b30e45569ac1283846f76f27f7bb5e908ee6deed57e88",
   "items_total": 20000,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 20000,
   "n_label1": 10000,
   "n_label0": 10000,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.9539,
   "accuracy_ci95": [
    0.9509,
    0.9568
   ],
   "balanced_accuracy": 0.9539,
   "auroc": 0.9796,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.979,
    "words": 0.974,
    "length": 0.972,
    "shape": 0.975,
    "alpha_first": 0.637,
    "alpha_last": 0.509
   },
   "headline": "surface baseline: 95.4% held-out accuracy [95.1%, 95.7%] from six probes, against 50.0% for the majority class; balanced accuracy 95.4%, AUROC 0.980."
  },
  {
   "name": "halueval/dialogue",
   "label_kind": "right_vs_wrong_answer",
   "status": "scored",
   "sha256": "9c461df2691e4362837f66fceaaff3bc260453350c3a1cae76a5a52e5e338bfd",
   "items_total": 20000,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 20000,
   "n_label1": 10000,
   "n_label0": 10000,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.7825,
   "accuracy_ci95": [
    0.7767,
    0.7882
   ],
   "balanced_accuracy": 0.7825,
   "auroc": 0.8615,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.846,
    "words": 0.849,
    "length": 0.722,
    "shape": 0.603,
    "alpha_first": 0.56,
    "alpha_last": 0.512
   },
   "headline": "surface baseline: 78.2% held-out accuracy [77.7%, 78.8%] from six probes, against 50.0% for the majority class; balanced accuracy 78.2%, AUROC 0.862."
  },
  {
   "name": "aperture/knowledge-prominence-v1",
   "label_kind": "known_vs_unknown_real_entity",
   "status": "scored",
   "sha256": "955778ddef49b4c6fd3e330e0704add067413475772b082c55816048f5e6aeea",
   "items_total": 324,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 324,
   "n_label1": 162,
   "n_label0": 162,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 5,
   "accuracy": 0.4877,
   "accuracy_ci95": [
    0.432,
    0.5435
   ],
   "balanced_accuracy": 0.4877,
   "auroc": 0.4696,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.532,
    "words": 0.521,
    "length": 0.554,
    "shape": 0.516,
    "alpha_first": 0.528,
    "alpha_last": 0.518
   },
   "headline": "surface baseline: 48.8% held-out accuracy [43.2%, 54.4%] from six probes, against 50.0% for the majority class and 50.0% for knowing each stratum's label rate; balanced accuracy 48.8%, AUROC 0.470. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "aperture/knowledge-prominence-v2",
   "label_kind": "known_vs_unknown_real_entity",
   "status": "scored",
   "sha256": "be110cd56e479a469706f46f8dcd5fee4e8814d841c3e686cf30819796c5d49f",
   "items_total": 296,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 296,
   "n_label1": 148,
   "n_label0": 148,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 5,
   "accuracy": 0.5135,
   "accuracy_ci95": [
    0.455,
    0.5718
   ],
   "balanced_accuracy": 0.5135,
   "auroc": 0.5185,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.606,
    "words": 0.514,
    "length": 0.561,
    "shape": 0.504,
    "alpha_first": 0.514,
    "alpha_last": 0.511
   },
   "headline": "surface baseline: 51.3% held-out accuracy [45.5%, 57.2%] from six probes, against 50.0% for the majority class and 50.0% for knowing each stratum's label rate; balanced accuracy 51.3%, AUROC 0.518."
  },
  {
   "name": "aperture/calibration-battery-v1",
   "label_kind": "real_vs_fabricated_entity",
   "status": "scored",
   "sha256": "a31e55da3dc94f317f8e1886fcec47c6c60de4f9492e95c11aab1ed7297069d0",
   "items_total": 168,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 168,
   "n_label1": 84,
   "n_label0": 84,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 6,
   "accuracy": 0.9226,
   "accuracy_ci95": [
    0.8713,
    0.9582
   ],
   "balanced_accuracy": 0.9226,
   "auroc": 0.9803,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.976,
    "words": 0.748,
    "length": 0.851,
    "shape": 0.852,
    "alpha_first": 0.641,
    "alpha_last": 0.545
   },
   "headline": "surface baseline: 92.3% held-out accuracy [87.1%, 95.8%] from six probes, against 50.0% for the majority class and 50.0% for knowing each stratum's label rate; balanced accuracy 92.3%, AUROC 0.980."
  },
  {
   "name": "halueval/summarization",
   "label_kind": "right_vs_wrong_answer",
   "status": "scored",
   "sha256": "86d2561eeb4271eb50b6bd90c9fd38636d30eaf4e98b648045f221adb1c2c758",
   "items_total": 20000,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 20000,
   "n_label1": 10000,
   "n_label0": 10000,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.8719,
   "accuracy_ci95": [
    0.8672,
    0.8765
   ],
   "balanced_accuracy": 0.8719,
   "auroc": 0.9446,
   "majority_accuracy": 0.5,
   "stratum_prior_accuracy": 0.5,
   "probe_auroc_held_out": {
    "lexicality": 0.909,
    "words": 0.923,
    "length": 0.75,
    "shape": 0.593,
    "alpha_first": 0.545,
    "alpha_last": 0.51
   },
   "headline": "surface baseline: 87.2% held-out accuracy [86.7%, 87.6%] from six probes, against 50.0% for the majority class; balanced accuracy 87.2%, AUROC 0.945."
  },
  {
   "name": "halueval/general",
   "label_kind": "hallucinated_vs_faithful_response",
   "status": "scored",
   "sha256": "73302cd97666b808fb94b5d8a4a2c924049c9484e609017a2f808c4a6d539219",
   "items_total": 4507,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 4507,
   "n_label1": 815,
   "n_label0": 3692,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.8185,
   "accuracy_ci95": [
    0.8069,
    0.8297
   ],
   "balanced_accuracy": 0.5197,
   "auroc": 0.6945,
   "majority_accuracy": 0.8192,
   "stratum_prior_accuracy": 0.8192,
   "probe_auroc_held_out": {
    "lexicality": 0.688,
    "words": 0.69,
    "length": 0.535,
    "shape": 0.624,
    "alpha_first": 0.542,
    "alpha_last": 0.522
   },
   "headline": "surface baseline: 81.8% held-out accuracy [80.7%, 83.0%] from six probes, against 81.9% for the majority class; balanced accuracy 52.0%, AUROC 0.695. The surface earns nothing beyond the class prior here."
  },
  {
   "name": "ragtruth/responses",
   "label_kind": "hallucinated_vs_faithful_response",
   "status": "scored",
   "sha256": "e4c2e4ac24fff676d8984cc61c35d791612fadc58015335d97dd632375e18073",
   "items_total": 17790,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 8000,
   "n_label1": 3428,
   "n_label0": 4572,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 3,
   "accuracy": 0.7077,
   "accuracy_ci95": [
    0.6976,
    0.7177
   ],
   "balanced_accuracy": 0.689,
   "auroc": 0.7536,
   "majority_accuracy": 0.5715,
   "stratum_prior_accuracy": 0.7007,
   "probe_auroc_held_out": {
    "lexicality": 0.742,
    "words": 0.735,
    "length": 0.663,
    "shape": 0.625,
    "alpha_first": 0.538,
    "alpha_last": 0.542
   },
   "headline": "surface baseline: 70.8% held-out accuracy [69.8%, 71.8%] from six probes, against 57.1% for the majority class and 70.1% for knowing each stratum's label rate; balanced accuracy 68.9%, AUROC 0.754."
  },
  {
   "name": "apollo/roleplaying",
   "label_kind": "honest_vs_deceptive_response",
   "status": "scored",
   "sha256": "b80332e391a00cfa8eba758723b63535351d0b71f803e2012426c4d97ef30390",
   "items_total": 287,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 287,
   "n_label1": 203,
   "n_label0": 84,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.7352,
   "accuracy_ci95": [
    0.6801,
    0.7853
   ],
   "balanced_accuracy": 0.5825,
   "auroc": 0.7131,
   "majority_accuracy": 0.7073,
   "stratum_prior_accuracy": 0.7073,
   "probe_auroc_held_out": {
    "lexicality": 0.744,
    "words": 0.746,
    "length": 0.567,
    "shape": 0.551,
    "alpha_first": 0.532,
    "alpha_last": 0.572
   },
   "headline": "surface baseline: 73.5% held-out accuracy [68.0%, 78.5%] from six probes, against 70.7% for the majority class; balanced accuracy 58.2%, AUROC 0.713."
  },
  {
   "name": "apollo/insider-trading-report",
   "label_kind": "honest_vs_deceptive_response",
   "status": "scored",
   "sha256": "73bfa9f8da734d7c435a0641f4e5f464a5c4401b4d102b1f8c152174f90ffd44",
   "items_total": 1005,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 1005,
   "n_label1": 336,
   "n_label0": 669,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.996,
   "accuracy_ci95": [
    0.9898,
    0.9989
   ],
   "balanced_accuracy": 0.9948,
   "auroc": 0.9988,
   "majority_accuracy": 0.6657,
   "stratum_prior_accuracy": 0.6657,
   "probe_auroc_held_out": {
    "lexicality": 0.999,
    "words": 0.999,
    "length": 0.738,
    "shape": 0.502,
    "alpha_first": 0.5,
    "alpha_last": 0.694
   },
   "headline": "surface baseline: 99.6% held-out accuracy [99.0%, 99.9%] from six probes, against 66.6% for the majority class; balanced accuracy 99.5%, AUROC 0.999."
  },
  {
   "name": "apollo/insider-trading-doubledown",
   "label_kind": "honest_vs_deceptive_response",
   "status": "scored",
   "sha256": "73bfa9f8da734d7c435a0641f4e5f464a5c4401b4d102b1f8c152174f90ffd44",
   "items_total": 655,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 655,
   "n_label1": 631,
   "n_label0": 24,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 1,
   "accuracy": 0.974,
   "accuracy_ci95": [
    0.9588,
    0.9848
   ],
   "balanced_accuracy": 0.726,
   "auroc": 0.9804,
   "majority_accuracy": 0.9634,
   "stratum_prior_accuracy": 0.9634,
   "probe_auroc_held_out": {
    "lexicality": 0.982,
    "words": 0.982,
    "length": 0.852,
    "shape": 0.649,
    "alpha_first": 0.5,
    "alpha_last": 0.667
   },
   "headline": "surface baseline: 97.4% held-out accuracy [95.9%, 98.5%] from six probes, against 96.3% for the majority class; balanced accuracy 72.6%, AUROC 0.980."
  },
  {
   "name": "apollo/sandbagging",
   "label_kind": "honest_vs_deceptive_response",
   "status": "scored",
   "sha256": "2f57cd2abe2aeb3f47392e6212c982472840591c510118e8149383ec1dcb069c",
   "items_total": 932,
   "tool": "surface_baseline",
   "version": "0.1",
   "nulltest_version": "1.4",
   "n": 932,
   "n_label1": 420,
   "n_label0": 512,
   "folds": 5,
   "seed": 20260724,
   "features": [
    "lexicality",
    "words",
    "length",
    "shape",
    "alpha_first",
    "alpha_last"
   ],
   "n_strata": 2,
   "accuracy": 0.9936,
   "accuracy_ci95": [
    0.986,
    0.9976
   ],
   "balanced_accuracy": 0.9937,
   "auroc": 0.9998,
   "majority_accuracy": 0.5494,
   "stratum_prior_accuracy": 0.9356,
   "probe_auroc_held_out": {
    "lexicality": 0.999,
    "words": 1.0,
    "length": 0.893,
    "shape": 0.681,
    "alpha_first": 0.5,
    "alpha_last": 0.508
   },
   "headline": "surface baseline: 99.4% held-out accuracy [98.6%, 99.8%] from six probes, against 54.9% for the majority class and 93.6% for knowing each stratum's label rate; balanced accuracy 99.4%, AUROC 1.000."
  }
 ]
}