BN_MIN_STD = 1e-4 caps the per-unit gain at 1/1e-4 = 1e4, and the comment above
it states that as though it were a safety property. It is not. A unit whose
running variance is ~0 is a CONSTANT feature carrying no information, and
dividing its rounding noise by 1e-4 hands the next layer an activation of
several hundred. BN's contract is "output has ~unit variance"; a unit that
cannot supply that must contribute nothing, not the largest signal in the layer.
MEASURED, 2026-08-17 SP500 H4, four topologies on identical separate charts:
model spread Neutral CHOSE Neutral TIED rail
CONV 0.386 0.68% 0.10% 0.48%
LSTM 0.392 0.63% 0.00% 0.00%
HYB 0.376 1.79% 0.00% 0.01%
PAI 0.192 0.09% 80.63% 99.99%
bn1's cached nx normed 1.38e4 over 800 units. PAI's SIGMOID head was on its
rails on 99.99% of bars, with Buy and Sell landing on the SAME rail so they
compared exactly equal, and ApplyClassificationSoftmax()'s strict-majority rule
reported that tie as Neutral on ~80% of bars.
So the long-running "PAI is heavily biased toward Neutral" was never a
class-prior problem: the net CHOSE Neutral on 0.09% of bars. It was float
equality on a saturated head. The 331ab29 counters answered it on their first
run.
PAI-only because it is the one topology whose FIRST batch norm sits on the raw
800-dim input vector - CONV/LSTM/CONVLSTM all have a conv or LSTM stage in
front, so their first BN sees a learned representation with no degenerate
units. That asymmetry was already on file as a suspicion; this is the mechanism.
FIX, mirrored in both backends (host NeuronBatchNorm.mqh and device Network.cl):
forward nx = clamp(delta/sd, -BN_MAX_NX, +BN_MAX_NX), BN_MAX_NX = 8
backward if the forward bound this unit, the output stopped depending on the
input, so d(nx)/dx = 0 and NO gradient passes
The backward half is not optional. g is divided by the same sd the forward
multiplies by, so a degenerate unit gets its GRADIENT amplified 1e4x too - the
"receives gradients divided by sqrt(var) ~ 500" pathology already noted in
Network.cl's Adam kernel. Bounding only the forward would move the explosion
downstream.
8 sigma is inert on anything healthy (|nx| > 8 is a ~1e-15 event under
normality); it binds only on degenerate units, which is the entire point. Same
clamp-to-range idiom the activation derivatives beside it already use.
SelfCheckBnForward/SelfCheckBnHiddenGrad already prove host against kernel, and
BN_OPT_NX was already consumed in the backward for the gamma gradient, so the
new read adds no lifetime assumption.
Expect PAI to change behaviour and CONV/LSTM/CONVLSTM not to (their rail rate is
~0%, so the clamp never binds). No .nnw format or fingerprint change.
ALSO: print the zero-skill reference on the era line. m_oosWinLongTotal and
m_oosWinShortTotal have been accumulated for a long time and NEVER printed,
which is why three separate topologies all sitting at 62% read as a mysterious
coincidence rather than the obvious base rate. It is not a coincidence: with the
target (1.70 ATR) nearer than the stop (3.07 ATR), BOTH sides win on 24.5% of
bars, so winLong+winShort covers ~124% of them and a no-edge caller collects
62.1% whichever way it calls - against a 64.3% break-even. Derived from this
run's own label counts: (11329 - 2776 + 2*2776) / (2*11350) = 62.14%.
Every win rate on that line must be read against this, not against 50%.
NOT COMPILED - user compiles.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
d30420e3f2