Prior-doubt experiment — objective vs LLM-judge labels
Hypothesis: perturbation features only 'help' on LLM-judge-disagreement subsets

label variant                                         n    pos      A4      A1      A2    Δ(A4−A1)
TriviaQA (OBJECTIVE, string-match)                  200  0.280  0.7457  0.7517  0.6559     -0.0061
NQ-Open (OBJECTIVE, string-match)                    75  0.547  0.4240  0.4907  0.4096     -0.0667
TruthfulQA (SINGLE LLM judge)                       200  0.635  0.6915  0.7183  0.6011     -0.0268
TruthfulQA (MAJORITY 3 LLM judges)                  200  0.560  0.6607  0.6417  0.6411     +0.0190
TruthfulQA UNANIMOUS (easy cases)                   140  0.571  0.6746  0.6965  0.6073     -0.0219
TruthfulQA DISAGREE subset — single label            23  0.826  0.6842  0.6974  0.5395     -0.0132
TruthfulQA DISAGREE subset — majority label          23  0.174  0.6842  0.6974  0.5395     -0.0132

== Key contrasts ==
  TriviaQA  Δ = -0.0061  (objective ground truth)
  NQ-Open   Δ = -0.0667  (objective ground truth)
  TruthfulQA single   Δ = -0.0268  (single LLM judge)
  TruthfulQA majority Δ = +0.0190  (3 LLM judges vote)
  TruthfulQA unanimous Δ = -0.0219  (easy: 3 judges agree)
  TruthfulQA disagree subset (n=23):
     with single label: Δ = -0.0132
     with majority label: Δ = -0.0132

== Interpretation rules ==
  If Δ > 0 only on LLM-consensus labels, perturbations track LLM-disagreement, not truth.
  If Δ ≤ 0 on objective labels AND Δ > 0 on majority-LLM label, yuragi is a judge-shift detector.
  If Δ flips sign on the 'disagree subset' itself, the story is fully coherent.
