{
  "design": "2x2 (four checkpoints) -- the 2x3 was never trained",
  "cbm1_noise_floor": {
    "per_condition": {
      "C1-vanilla-general": {
        "observed": 0.0008721585366141058,
        "noise_floor": 0.0004414020201011664,
        "noise_fraction": 0.5061029635905159,
        "implied_true_between_variance": 0.00043075651651293943,
        "below_own_noise_floor": false
      },
      "C2-vanilla-dharma": {
        "observed": 0.0007137163330472936,
        "noise_floor": 0.0004623234986495341,
        "noise_fraction": 0.6477692568356829,
        "implied_true_between_variance": 0.0002513928343977595,
        "below_own_noise_floor": false
      },
      "C3-contemplative-general": {
        "observed": 0.0007737611049368642,
        "noise_floor": 0.00022990935463150904,
        "noise_fraction": 0.29713221970529097,
        "implied_true_between_variance": 0.0005438517503053551,
        "below_own_noise_floor": false
      },
      "C4-contemplative-dharma": {
        "observed": 0.0001537419865178794,
        "noise_floor": 0.00040965080092381247,
        "noise_fraction": 2.6645343292489083,
        "implied_true_between_variance": -0.0002559088144059331,
        "below_own_noise_floor": true
      }
    },
    "n_prompts_per_domain": 12,
    "design_target_per_domain": 100,
    "conditions_below_noise_floor": [
      "C4-contemplative-dharma"
    ],
    "verdict": "At n=12 prompts/domain the sampling-noise term is the same order as the measured score itself. Below its own noise floor: C4-contemplative-dharma -- implied true between-domain variance is negative, i.e. indistinguishable from a model with identical confidence across all five domains. Any SC1 verdict resting on that condition is not yet supported by the data."
  },
  "cbm1_sc1_bootstrap": {
    "point_estimate": 0.4151473680563619,
    "ci95": [
      -1.1845641650524654,
      0.8176405991130331
    ],
    "median": 0.3716717997972229,
    "p_meets_sc1_target": 0.5698,
    "p_positive": 0.7773,
    "n_boot": 10000,
    "resampling_unit": "domain mean (normal approximation, n=12 prompts/domain)",
    "caveat": "Parametric, not nonparametric -- per-prompt max-probs are not banked. Assumes each domain mean is normally distributed; max-softmax is bounded in [0,1] and right-skewed, so the true interval is likely mildly asymmetric."
  },
  "cbm1_sc1_bootstrap_nonparametric": {
    "point_estimate": 0.4151473680563619,
    "ci95": [
      -1.1043726292201472,
      0.8168258668261281
    ],
    "median": 0.3612885911446887,
    "p_meets_sc1_target": 0.5627,
    "p_positive": 0.7773,
    "n_boot": 10000,
    "n_prompts_per_domain": 12,
    "resampling_unit": "prompt (nonparametric, resampled within domain)"
  },
  "cbm1_condition_level_permutation": {
    "observed_abs_reduction": 0.4151473680563619,
    "n_labellings": 6,
    "n_distinct_partitions": 3,
    "p_exact": 0.8333333333333334,
    "min_attainable_p": 0.3333333333333333,
    "significant_at_05": false,
    "verdict": "STRUCTURALLY UNDERPOWERED. With 4 checkpoints there are only 3 distinct 2-vs-2 partitions, so the smallest p this test can ever return is 0.333 -- above alpha=0.05 before any data is seen. Condition-level permutation cannot support the \u00a74.4 claim at any effect size. Use replicated units (domains/layers/prompts)."
  },
  "cbm3_paired_layer_permutation": {
    "per_layer_delta": [
      0.018167142570506112,
      0.07280596615768609,
      0.058316099598207494,
      0.07158561956401477,
      0.052269953332686026,
      0.09039739882785636
    ],
    "mean_delta_nats": 0.06059036334182614,
    "ci95": [
      0.041573122200791124,
      0.07677172697294503
    ],
    "p_two_sided_exact": 0.03125,
    "n_sign_assignments": 64,
    "min_attainable_p": 0.03125,
    "cohens_d_paired": 2.4602164752487563,
    "all_layers_same_sign": true,
    "significant_at_05": true,
    "resampling_unit": "transformer layer (n=6), paired within layer",
    "caveat": "Layer is the only replicated unit banked -- cbm3.json stores per-layer aggregates over 60 inputs, not per-input entropies. Layers within a model are not independent, so this p is optimistic; treat it as a consistency check on direction, which is where CBM-3's strength actually lies."
  },
  "bonferroni": {
    "family_size": 2,
    "alpha": 0.05,
    "corrected_threshold": 0.025,
    "survivors": [],
    "p_values": {
      "cbm1_sc1_condition_permutation": 0.8333333333333334,
      "cbm3_entropy_paired_layer_permutation": 0.03125
    }
  },
  "release_note": "Historical CBM1/3 analysis. CBM2/4 were subsequently measured September 14. Bayes factor not computed because no prior was specified before inspection. Layer and condition permutations do not establish training-seed generalization."
}