{
 "gpt-6-astra": {
  "means": {
   "bare": 1.7995452380952381,
   "show_all": 1.6257452380952382,
   "bare2": 2.8301738095238096,
   "attempt_then_consult": 2.5481023809523813,
   "consult_on_failure": 2.854271428571429,
   "audit_then_consult": 2.483002380952381,
   "prior_then_consult": 2.9066238095238095
  },
  "contrasts": {
   "PRIMARY attempt_then_consult - bare2": {
    "mean": -0.2820714285714286,
    "lo": -0.6701214285714285,
    "hi": 0.05375714285714292,
    "up": 5,
    "down": 14
   },
   "audit_then_consult - bare2": {
    "mean": -0.3471714285714286,
    "lo": -0.8020619047619048,
    "hi": 0.010947619047619069,
    "up": 4,
    "down": 16
   },
   "consult_on_failure - bare2": {
    "mean": 0.02409761904761902,
    "lo": -0.07458809523809524,
    "hi": 0.15687380952380947,
    "up": 2,
    "down": 2
   },
   "attempt_then_consult - prior_then_consult": {
    "mean": -0.35852142857142855,
    "lo": -0.6251547619047619,
    "hi": -0.13984999999999986,
    "up": 4,
    "down": 15
   },
   "audit_then_consult - attempt_then_consult": {
    "mean": -0.06510000000000002,
    "lo": -0.3120904761904762,
    "hi": 0.1992714285714286,
    "up": 9,
    "down": 11
   },
   "prior_then_consult - bare2": {
    "mean": 0.07645000000000003,
    "lo": -0.162402380952381,
    "hi": 0.3197,
    "up": 12,
    "down": 8
   },
   "show_all - bare": {
    "mean": -0.17380000000000004,
    "lo": -0.482,
    "hi": 0.07590238095238087,
    "up": 9,
    "down": 12
   },
   "bare2 - bare": {
    "mean": 1.0306285714285714,
    "lo": 0.5572571428571428,
    "hi": 1.5511309523809527,
    "up": 17,
    "down": 1
   },
   "attempt_then_consult - bare": {
    "mean": 0.7485571428571428,
    "lo": 0.37105714285714275,
    "hi": 1.2009000000000003,
    "up": 15,
    "down": 3
   },
   "audit_then_consult - bare": {
    "mean": 0.6834571428571429,
    "lo": 0.27835714285714286,
    "hi": 1.1519142857142857,
    "up": 14,
    "down": 3
   },
   "consult_on_failure - bare": {
    "mean": 1.0547261904761904,
    "lo": 0.5724238095238094,
    "hi": 1.5919285714285716,
    "up": 18,
    "down": 0
   },
   "prior_then_consult - bare": {
    "mean": 1.1070785714285716,
    "lo": 0.6324595238095239,
    "hi": 1.62975,
    "up": 18,
    "down": 1
   },
   "attempt_then_consult - show_all": {
    "mean": 0.922357142857143,
    "lo": 0.4816666666666667,
    "hi": 1.422809523809524,
    "up": 15,
    "down": 6
   },
   "consult_on_failure - prior_then_consult": {
    "mean": -0.05235238095238101,
    "lo": -0.31961666666666677,
    "hi": 0.2061357142857141,
    "up": 8,
    "down": 12
   },
   "audit_then_consult - prior_then_consult": {
    "mean": -0.4236214285714286,
    "lo": -0.8015071428571429,
    "hi": -0.11070000000000002,
    "up": 4,
    "down": 16
   }
  },
  "diag": {
   "P2_consult - P2_refine (raw second pass)": {
    "mean": -0.24386428571428573,
    "lo": -0.6701738095238096,
    "hi": 0.14820714285714287
   },
   "P2_audit - P2_refine (raw second pass)": {
    "mean": -0.3374404761904762,
    "lo": -0.8102309523809523,
    "hi": 0.06374285714285713
   },
   "P2_prior - P2_refine (raw second pass)": {
    "mean": -0.06270238095238088,
    "lo": -0.3745119047619045,
    "hi": 0.2287571428571428
   },
   "S - P1 (raw pass evals)": {
    "mean": -0.19142380952380955,
    "lo": -0.5005761904761905,
    "hi": 0.06092857142857138
   }
  }
 }
}