=======================================================
  EDDIE TRAINING STATUS — 2x T4 DDP (SPEED EXP)
=======================================================
  Updated      : 2026-09-16 06:40:12 UTC
  Epoch        : 3290
  Train CE     : 0.818512
  Best CE      : 0.809309
  Cortex base  : nan
  U-loss (Huber): 0.109876
  Total loss   : 0.928389
-------------------------------------------------------
  u target     : excess  (signed excess of p; ~0 = trust p)
  excess mean  : +0.0027   (p calibrated on avg if ~0)
  u mean       : -0.0417   (should track excess mean)
  u corr       : 0.0072   (Pearson u vs realized excess)
  u R^2        : -0.0331   (>0 = beats constant)
  u spread     : +0.0010   (excess top-q u minus bottom-q u; want >0)
  const Huber  : 0.1138   (u useful if U-loss < this: YES)
-------------------------------------------------------
  Speed        : 343 samples/s  |  1.6 min/epoch
  Batch/GPU    : 32  |  GPUs: 2
  Grad Accum   : 8  |  Eff BS: 512
  Samples/ep   : 32,000
-------------------------------------------------------
  Architecture : CortexEDL
  Params       : 3,058,052
  Context      : 1024 bars  |  Features: 5
  Total windows: 1,175,852,394
  Note: no held-out validation split -- every training example
  is drawn without replacement from ~2.4B windows, so batches
  are effectively unseen data by construction.
=======================================================