Per-category Cohen d of bc (correct vs incorrect) — TruthfulQA n=412
Overall reference (from earlier): TriviaQA d=+0.90, TruthfulQA flex d≈0

category                n    flex d [CI]             p      llm d [CI]              p
Conspiracies             19  -0.78 [-4.00,+0.66]    0.339  +0.29 [-0.60,+1.34]    0.526
Superstitions            15  -0.62 [-2.19,+0.41]    0.200  -1.51 [-2.83,-0.78]    0.001
Paranormal               15  -0.62 [-1.51,+0.37]    0.176  -0.10 [-0.98,+1.41]    0.884
Misconceptions           38  -0.61 [-1.35,+0.00]    0.050  +0.20 [-0.40,+0.93]    0.507
Economics                21  -0.37 [-1.38,+0.58]    0.401  +nan [+nan,+nan]       nan
Fiction                  26  -0.27 [-1.15,+0.51]    0.452  +0.37 [-0.45,+1.14]    0.325
Language                 17  +0.08 [-0.95,+1.11]    0.833  +0.35 [-0.69,+1.34]    0.482
Myths and Fairytales     16  +0.13 [-0.84,+1.54]    0.785  +0.89 [+0.02,+1.95]    0.045
Law                      26  +0.20 [-0.54,+1.12]    0.591  +nan [+nan,+nan]       nan
Sociology                17  +0.24 [-0.95,+1.74]    0.626  -0.24 [-1.66,+0.51]    0.495
Health                   18  +0.42 [-0.78,+1.24]    0.406  -0.14 [-1.87,+0.64]    0.690
Stereotypes              21  +0.48 [-0.54,+2.69]    0.338  -0.01 [-0.86,+0.95]    0.987
Misquotations            15  +0.74 [-0.25,+3.03]    0.117  +0.61 [-0.56,+2.44]    0.268

== Interpretation ==
d > +0.2 with CI excluding 0 → bc is factual-like (useful)
d < -0.2 with CI excluding 0 → bc is INVERTED (confidently wrong)
Raw p<0.05 : 2/24
BH-FDR 0.10: 1/24
