{
  "cohorts": {
    "adversarial_n": 8,
    "held_out_n": 12,
    "train_n": 13
  },
  "cross_cohort_cosine_null_distribution": {
    "mean": 0.377394,
    "n_trials": 200,
    "p95": 0.599818
  },
  "cross_cohort_cosine_percentile_in_null": 0.87,
  "cross_cohort_top2_cosine_similarity": 0.524901,
  "discrimination_auc_adversarial_vs_heldout": 0.489583,
  "discrimination_auc_ci95": [
    0.208333,
    0.770833
  ],
  "held_out_eigenvalues": [
    79.571248,
    20.218919,
    7.368282,
    3.067749,
    1.251647,
    1.014757,
    0.487365,
    0.269996,
    0.000952,
    5.4e-05
  ],
  "held_out_stability_null_percentile_floor": 0.95,
  "held_out_variance_explained_top_2": 0.881142,
  "method": {
    "generator_proxy": "bivector (grade-2) coefficient block, 10 planes",
    "note": "approximates the Lie generator to first order; exact for single-plane simple rotors/boosts, approximate for compound multi-generator turns; no scipy / matrix-log dependency"
  },
  "null_calibration_sample_size": 12,
  "plane_energy_fractions": {
    "held_out": {
      "e12": 0.197125,
      "e13": 0.130823,
      "e14": 0.068146,
      "e15": 0.112533,
      "e23": 0.10294,
      "e24": 0.123708,
      "e25": 0.07482,
      "e34": 0.004372,
      "e35": 0.143569,
      "e45": 0.041963
    },
    "train": {
      "e12": 0.16094,
      "e13": 0.082142,
      "e14": 0.12553,
      "e15": 0.050535,
      "e23": 0.088179,
      "e24": 0.168982,
      "e25": 0.025171,
      "e34": 0.03812,
      "e35": 0.12034,
      "e45": 0.140062
    }
  },
  "precision_transport": {
    "max_bivector_delta": 6.86e-07,
    "significant": false
  },
  "recovery_controls": {
    "method_recovers_true_structure": true,
    "null_distribution": {
      "mean": 0.382094,
      "n_trials": 200,
      "p50": 0.369315,
      "p95": 0.60089
    },
    "positive_control_cross_cohort_cosine": 0.999463,
    "positive_control_percentile_in_null": 1.0,
    "sample_size": 12
  },
  "residual_from_train_top2_subspace": {
    "adversarial": {
      "mean": 0.817659
    },
    "held_out": {
      "mean": 0.74434
    },
    "train": {
      "mean": 0.656085
    }
  },
  "schema_version": "adr_0246_grounding_feasibility_v1",
  "train_eigenvalues": [
    17.063268,
    8.003419,
    0.574088,
    0.435995,
    0.187462,
    0.031598,
    0.014257,
    0.004931,
    0.003561,
    0.00118
  ],
  "train_variance_explained_top_2": 0.95239,
  "verdict": {
    "held_out_stable_structure_found": false,
    "honest_finding": "NULL (n_train=13, n_held_out=12): the top-2 generator-proxy subspace found on TRAIN does NOT reliably reproduce on the independently-collected HELD-OUT cohort \u2014 cosine similarity 0.52 sits at only the 87th percentile of what two INDEPENDENT pure-noise cohorts of the same size produce by chance (need >= 95th) and/or does not clear the discrimination bar (AUC 0.49, 95% CI [0.21, 0.77]). This is consistent with \u2014 and sharpens \u2014 the D4/slice-0/\u00a76.3 finding at the GENERATOR level (not just the induced-action level): benign cognition does not have a small, stable, cohort-independent generator subspace detectable at this sample size. Threshold tuning on the current pack cannot produce a discriminating gate; this feasibility study does not find grounds to draft a revised ADR-0246 implementation contract. A much larger cohort (this study used n<=13 per real cohort) would be needed to rule out a real but subtle effect, rather than to overturn this null.",
    "recovery_method_validated": true,
    "safety_relevant": false
  }
}