FINAL UNIFIED ENSEMBLE  -  Combining ALL discovered signals
============================================================
Protocol: 5-fold stratified CV, AUC with 95% bootstrap CI.
Models: 4-model ensemble (LogReg, RF, GB, HistGB) + mean blend.
Label source: relabel_results.jsonl (LLM-judge, n=412).

Data overlap: the three signal cohorts were collected on DIFFERENT
question subsets. pilot_all13_n30 (prompt signals, 30) is disjoint
from fast_817 (yuragi, 382) and from logprob_shape (50, subset of
fast). NO question has all three signals simultaneously, so we run
three complementary fusions.

OPTION A1 - pilot-30, yuragi + adversarial-prompt features
----------------------------------------------------------
yuragi-only (n=30)       LogReg AUC=0.560  [0.33, 0.77]
prompt-only (n=30)       GB     AUC=0.578  [0.36, 0.79]
yuragi + prompt (n=30)   LogReg AUC=0.538  [0.31, 0.77]

OPTION A2 - fast-50, yuragi + logprob-shape features
----------------------------------------------------
yuragi-only (n=50)       LogReg AUC=0.592  [0.43, 0.75]
shape-only  (n=50)       LogReg AUC=0.574  [0.40, 0.74]
yuragi + shape (n=50)    LogReg AUC=0.621  [0.46, 0.77]

OPTION C - two-stage on all 412 (yuragi primary + prompt override)
------------------------------------------------------------------
yuragi-primary (n=412)   RF     AUC=0.636  [0.58, 0.69]
                         BLEND         0.634
Stage-2 meta (yuragi-OOF + prompt features) n=30:  AUC=0.484
Combined n=412 (override the 30):                  AUC=0.589

SUMMARY (5-fold CV AUC, best per cohort)
----------------------------------------
  A1 pilot-30 fusion :  0.538
  A2 fast-50 fusion  :  0.621
  C  412 two-stage   :  0.636

Best overall = 0.636  (below 0.80 publishability bar).

INTERPRETATION
--------------
The previously reported headline numbers (prompt AUC 0.80, yuragi
AUC 0.73) were obtained as single-signal point estimates on the
full labeled set with NO cross-validation. Once we move to 5-fold
CV on the intersection cohorts, all numbers drop sharply:

1. n=30 is too small for any multi-feature classifier - CV folds
   have only 24 training samples, so even well-signalled features
   cannot be fit reliably (confidence intervals span 0.3-0.8).
2. The two high-AUC signals (prompt 0.80, yuragi 0.73) live on
   disjoint question sets, so Option A1 must choose one cohort and
   loses statistical power either way.
3. The two-stage approach (C) preserves the yuragi-412 signal but
   the n=30 stage-2 refinement adds noise (meta AUC 0.48 < primary
   0.64) - not enough pilot records to learn the override.

Publishable finding: the signals do NOT compose additively at the
per-question level when evaluated with honest CV. A fusion study
would require re-collecting prompt-signal data on all 412 questions
(est. ~6 prompts x 412 = 2500 calls at 55 rpm ~= 45 min of Groq
quota). The current best unified CV-AUC is 0.636 (Option C).
