Three backends left, as the operator specified: OpenCL, the CPU DLL,
and pure MQL5. CDirectMLMy was a two-tier wrapper (GPU via
WarriorDML.dll, CPU via WarriorCPU.dll) whose name only ever named the
tier being removed here; the CPU DLL tier - the one actually used on
the training machine (no OpenCL, no DirectML) - is untouched.
AI/NeuronDirectML.mqh -> AI/ComputeDll.mqh: dropped the DML_* #import
block and COMPUTE_TIER_GPU (checked first that nothing persists the
enum value and only one external site reads .Tier() - safe), collapsed
every tier==CPU?CPU_x():DML_x() ternary to a straight CPU_x() call.
Renamed CDirectMLMy->CComputeDll, InitDirectML()->InitComputeDll(),
member directml/DirectML->computeDll/ComputeDll across every AI/ file
that touched a neuron/net backend plus Topology.mqh/OnlineLearning.mqh.
NetBuild.mqh's InitComputeDll also lost the dead D3D12 error-code
switch and the now-impossible GPU-tier log branch.
Verified via per-file brace-balance diff against HEAD and a whole-repo
grep for every removed symbol (CDirectMLMy/InitDirectML/
COMPUTE_TIER_GPU/DML_*) - the only surviving hit is an intentional
historical-note comment in the new file's header.
DirectML\WarriorDML.cpp/.h and its build scripts are now orphaned C++
source, left in place pending an operator decision. Architecture docs
(AI_NETWORK.md, Warrior_EA_System_Overview.md, etc.) still describe the
4-backend/GPU-tier shape and are not updated in this pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The //| box blocks were excluded from 0b06f8e and 5efdb48 and were what
remained: 160 of them ran to 10+ lines, the longest to 88. Compressed to their
leading topic sentences - 5 lines for a function header, 8 for a file header -
keeping the box format and the standard MQL5 name/author lines verbatim.
Verified at the BYTE level this time, across every in-scope file: the list of
non-comment lines is byte-identical to HEAD and braces balance. The first check
compared a locale-decoded 'git show' against a UTF-8 read and flagged 25 files
that had not changed at all - every BOM and every non-ASCII line mismatched.
47,696 -> 40,665 lines in scope; comment share 38% -> 26%.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Same pass as 0b06f8e, applied file by file: comment runs of 4+ lines compressed
to their leading topic sentences, capped at 4 lines, whole sentences only.
Warning sentences (NEVER / MUST / trap / would-have) survive the budget.
Every file was checked the same way before committing: the list of non-comment
lines is byte-identical to HEAD, and braces balance. No code was touched.
Panel/, Enumerations/ and the already-terse System headers needed little or
nothing - PooledGate, TradeChecks, BinomialStats and Random came through with
no blocks over the threshold at all.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
MetaEditor: "declaration of 'eta' hides global variable" (Math.mqh:792
vs Network.mqh:80). The standard library's Math\Stat\Math.mqh declares a
local `double eta` in its incomplete-gamma branch, and our bare global
of the same name is in scope there.
Same fault as the b1/b2/lr/momentum macros retired in ea2552e: a
single-token global name living in a header that library code gets
compiled beside. The library cannot move, so ours does - 112 references
across 13 files, whole-word only.
Named g_eta rather than g_learningRate to stay inside the vocabulary
already around it (ETA_DECAY_FACTOR, ETA_MIN, m_etaCeiling, etaBefore),
all of which are untouched and none of which shadow anything.
One log line said "continuing to explore without decaying eta", where
the word was prose rather than a symbol reference; that reads "the
learning rate" now instead of naming a variable at the trader.
Scanned for the next occurrence rather than waiting for it: the only
other bare lowercase globals in the tree are eaName and tableschema,
both distinctive enough not to collide with a library local.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The gate's NormalUpperTail was a hand-rolled Abramowitz & Stegun 26.2.17
approximation. Its own comment gave the reason - "drags a chain of headers
behind it" - and that turned out to be one file: Math\Stat\Normal.mqh
includes only Math.mqh, which includes nothing. Swapped for Cody's rational
approximation in the library (~18 significant digits vs |error| < 7.5e-8).
No past verdict changes: at the z the gate operates on, the difference is
orders of magnitude below DEPLOY_FAMILY_WISE_ALPHA.
Adopting it needed the four bare macros in AI\Network.mqh gone first.
"#define b1 AdamBeta1" collides with an identifier in Math.mqh, so the
include would have macro-expanded the library's own local and failed to
compile - the same landmine that made the original author rename the
approximation's coefficients to ntB1..ntB5 rather than use the reference's
b1..b5. lr, b2 and momentum are the same class of hazard: single-token
global macros in a 52k-line codebase. All four now resolve to the input
names they always aliased, which is a pure textual identity - verified zero
bare occurrences remain.
Also:
- SelectionSort over the buffered signals was O(n^2) with an O(n^2) count of
StructToTime calls, because the comparison rebuilt both datetimes from the
six int date fields every time. Now materialises the keys once and does an
insertion sort; ArraySort cannot permute a struct array. IsEarlier goes
with it, MakeDateTime becomes SignalTime.
- Seven FileOpen sites lacked FILE_SHARE_READ|FILE_SHARE_WRITE, including
AtomicWriteBegin, which stages every model save. All 43 sites now carry
them - an exclusive open fails outright when another process holds the
path, which here has meant a silently skipped save.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both 2026-08-19 reports were the same staleness: every source behind the
label was an ERA artifact (live cache refills at pass-3 completion, the
snapshot copies once per era, dPrevSignal is the frozen purge-band edge
bar) - so the readout stepped at era cadence at best, stayed glued to
one direction, and lagged the era counter.
DisplayInference(): throttled (4s, 1s across an era boundary),
SIDE-EFFECT-FREE forward of the current decision bar (window ending on
bar 1, same question the live path asks) through the LEARNER net.
Batch-norm running stats are bracketed frozen/RESTORED via the new
CNet::GetBatchNormFrozen() + CNeuronBatchNormOCL::StatsFrozen() - restore,
not unfreeze, because a display tick can land between pass-3 chunks whose
whole scan holds them frozen. Writes nothing a trading or training path
reads (dPrevSignal, NMS state, tallies, watermarks all untouched;
RefreshLatestSignal is not reusable here precisely because it writes all
of them). LSTM safe by construction: h/c zeroed per forward.
ProspectiveVote() reads the fresh forward as its FIRST source; the
era-artifact chain becomes the fallback (meta head, warm-up, window
holes).
DisplayHudLine(): the reference library's training label, per ensemble
member - name, output activations (softmax probs or raw scalar), the
decision, its weighted vote (the exact consensus numerator term), era,
recent average error, "(trn)" while not vote-capable. Rendered under the
vote line in RefreshVoteReadout BEFORE the live-vote defer (member lines
are telemetry, not tradable readings), coloured by the member's own
direction in muted tones - the vote line's strict
green-only-when-it-would-trade rule is untouched.
NOT COMPILED - user compiles in MetaEditor.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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>
Three changes, one theme: the trade placed, the trade graded, and the trade
computed are now the same trade.
1) GEOMETRY WIRE (correctness, the ranked #1 open issue). The measured barrier
pair reached the LABELS only - OpenParams still placed orders at the enum
geometry (2*ATR/6*ATR), so the deploy gate certified "reaches 1.62*ATR before
3.33*ATR above break-even" about trades the EA never placed. Published via
g_DerivedSlAtrMult/g_DerivedTpAtrMult (ConfidenceBridge, same same-tick
contract as the confidence globals, because OpenParams runs on the root signal
which has no pointer to the AI filter). Two writers: DeriveBarrierGeometry at
era 0, and the .cfg adoption a deployed model takes. Overrides both legs and
both Intelligent modes - the certificate is exact or it is nothing. TP is
ATR-anchored like the label, NOT risk-relative, so a floor-widened stop cannot
reshape the certified target.
2) BATCH NORM RUNS DEVICE-SIDE ON OPENCL. Four kernels in Network.cl -
forward, hidden gradient, gamma/beta accumulate, gamma/beta apply - each a
line-for-line transcription of the host implementation (NormalizeHost /
HiddenGradHost / StepGammaBeta) including every NaN guard, clamp, and the
exact moment-write ordering. The host copies remain the runtime for the DLL
and pure-MQL5 tiers and the reference the kernels must match.
Because this box has no OpenCL platform, the safety story is layered:
- shim validation: kernels compiled as C and driven against a fp64 host
transcription over NaN-poisoned stats, NaN gamma, over-clamp inputs, the
frozen path, both optimizers, 3 batches - ALL PASS, worst normalized diff
0.132 vs tolerance 1.0
- in-situ self-check: each kernel is compared against its host twin ON FIRST
USE on the real device (SelfCheckBn*), covering what the shim cannot - arg
indices and buffer bindings. Any disagreement resyncs from the good copy,
latches all BN kernels off process-wide, and training continues host-side.
A transcription bug costs a warning and some speed, never a poisoned .nnw.
- sync discipline: BatchOptions is now a CBufferDouble with explicit
authority tracking (m_bnDeviceAuthoritative). Checkpoints/saves pull
read-only; restores/loads/resets push; a mid-batch handover drains the
device gamma/beta accumulator into the host arrays so no sample is lost.
3) SMALL FIXES. Apply-kernel build failure now latches the dispatch path at
init (one warning instead of warning + failed Execute). Build tag bumped to
win-scoring-gpu-v1 - first tag change since expectancy-stop-v1 despite five
binary-changing commits.
Both build variants compile 0 errors, 0 warnings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Batch norm computes host-side while its neighbours are device-resident, so
every value it touches crosses the bus - and the cost is the BLOCKING SYNC per
crossing, not the bytes. Per sample it did four reads, two of which were exact
duplicates:
feedForward -> previous layer's Output
calcInputGradients -> own Gradient, and the previous layer's Output AGAIN
update/accumulate -> own Gradient AGAIN
Nothing writes the previous layer's Output between the forward and backward
passes, and nothing writes this layer's Gradient between the gradient pass and
the weight-update pass - backPropOCL runs those as two separate top-to-bottom
loops and only the first writes gradients. So the repeats are removable and
the result is bit-exact: the same values, read once instead of twice.
Each cache is armed by its producer and disarmed by feedForward, so a consumer
whose producer did not run this sample falls back to reading the buffer rather
than using the previous sample's data. That is not hypothetical - a batch norm
at layer 1 never gets calcInputGradients called at all, the same asymmetry
backPropOCL already documents for a layer-1 LSTM, so its gradient cache is
never armed and it takes the fallback every time.
Both build variants compile 0 errors, 0 warnings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause of the B=32 regression, and it predates F4 entirely. Every Adam
kernel stored v already square-rooted and then fed that stored value back in
as if it were the variance:
v_new = sqrt(b2 * v_old + (1 - b2) * g^2)
That recursion has a fixed point at v ~= b2 = 0.999 for ANY gradient below
unit scale, so the denominator stops tracking the gradient and Adam degrades
into plain SGD with lr = lt. Measured against the shipped WarriorCPU.dll
(batch_accum_check.cpp, TestOptimizerScaleInvariance), 4000 steps of a
constant gradient: 3285x less displacement at |g|=1e-5 than at |g|=1, where
a scale-invariant optimizer gives the same distance for both. After the fix
all six magnitudes read 1.199 and v tracks |g| exactly.
It hit conv/LSTM specifically because they sit behind a batch-norm with
running variance ~2.6e+05, so their gradients arrive divided by ~500 - deep
in the degraded regime - while the dense stack near the loss stayed in the
working one. In situ on SP500 H1: lstm1 dW/W 2.62/10.0/7.14% -> 0.024/0.022/
0.003%, conv1 decaying to 0.000% by era 30. NeuronBatchNorm.mqh already
squared v back for gamma/beta and its comment named the kernels as wrong,
which is exactly why gamma/beta kept training while the stages behind froze.
Persisted .nnw needs no migration - v keeps its std-dev meaning.
Also, the two ways F4 exposed it, both mine:
- No LR compensation for B fewer steps per era. sqrt(B) for adaptive methods
(Krizhevsky 2014; Granziol et al. 2022), applied once in
InitialEtaForOptimizer(). Linear scaling (Goyal et al. 2017) is for SGD.
- Plateau patience denominated in eras, so raising B made the ladder 32x more
impatient in its only unit. PAI converged at era 41 on ~49k updates where
the same config had been finding new bests at era 1028.
TrainPlateauPatienceEras() stretches it by the same sqrt(B).
TRAIN_BATCH_SIZE 32 -> 8 so the patience stretch stays affordable (8 -> 23
eras per stage, not 8 -> 45). Both helpers are identities at B=1.
Deploy gate: DEPLOY_MIN_SIDE_RECALL_PCT (10%) folded into tradeableOK. The
perceptron reported Sell:0% recall in all 41 eras, cleared the floor on Buy
alone at 36.6% vs 34% chance, deployed, and sprayed buy arrows. Folded into
the ranking key rather than checked at deploy time so a one-sided era cannot
become best-so-far in the first place.
Deinit: the arrow purge now runs BEFORE ExtPanel.Destroy(), an unbounded
CAppDialog teardown that sat ahead of it - the same ordering inversion the
rule there exists to prevent. CONV was force-terminated 4.8 s into OnDeinit
(vs ~1.1 s for the three that finished) having reached none of its cleanup,
so its arrows stayed on the chart. Steps are now timed in the log.
PurgeChart's verification rescan filtered on OBJ_ARROW, the same blind spot
as the bulk delete, so "persisted 10 ... cleared 0" passed silently. It now
walks every object type and reports the object counts when both are zero.
Both build variants compile 0 errors / 0 warnings; both DLLs rebuilt.
FORCES A RETRAIN (already forced by N1) and both DLLs must ship with the .ex5.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Completes the 2026-08-09 training audit. FORCES A RETRAIN of every
Wyckoff-enabled config (N1 re-keys the fingerprint), and BOTH DLLs must be
redeployed alongside the .ex5 - they carry new exports.
F4 - mini-batch accumulation, TRAIN_BATCH_SIZE=32. Training was pure online
SGD (one weight update per bar), which is the mechanical source of the
era-to-era whipsaw every downstream guard was built to cope with. The O(n^2)
outer product is native - AccumulateWeightGrad / AccumulateWeightGradConv /
AccumulateBufferInto in Network.cl, WarriorCPU and WarriorDML - while the
optimizer step is host-side MQL5 shared by all tiers (ApplyAccumToBlock), so
there is one Adam/SGD implementation instead of four that can drift.
- the LSTM needs no outer-product kernel (WeightsGradient already holds the
sample's full dW) but could NOT simply be left un-zeroed between samples:
CPU_LSTMSeqBackward/DML_LSTMSeqBackward memset it on entry. Hence a
separate accumulator plus an elementwise add.
- batch-norm gamma/beta accumulate in host arrays, not new BatchOptions
slots - BN_OPT_STRIDE is baked into every persisted .nnw.
- scoped to pass 2; online learning keeps immediate updates. Every save /
checkpoint / scoring boundary flushes, scaling by the real sample count.
- degrades to per-sample updates (one log line) on a tier that cannot
accumulate, so old devices and DLL-free builds are unaffected.
- verified offline: DirectML/batch_accum_check.cpp drives the real exports
against an independent reference; at B=1 the accumulator matches the
shipped unbatched kernel's own gradient to 1.1e-16. Math only - the
in-situ check remains the per-layer dW/W report on a real era.
F6 - ComputeFirstLayerWidth budgeted against the RAW input width even where a
conv/LSTM front end had already reduced it, so an LSTM's dense stack was
charged for 1,280 inputs when it receives 64. Confirmed from the deployed
.cfg files: CONV, LSTM and HYBRID were all pinned at the 16-unit floor. Now
budgeted against the front-end output and capped at it (never fan out), with
the derivation reordered so both stages settle first.
N1 - EventCode/EventPhase/StructuralPhase are signed categoricals packing
direction and Wyckoff stage into one scalar across a sign discontinuity. Split
into direction + [0,1] magnitude, the same convention the base OHLC block uses.
Information-preserving; 13 readings now occupy 16 inputs.
Compiled clean (0 errors, 0 warnings); both DLLs rebuilt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four of the six findings from research/training_pipeline_audit_2026-08-09.md
(F4 mini-batching and F6 feature re-encode deliberately deferred - see the
report's implementation-status section for why):
- F1: pass-2 Fisher-Yates (and AutoTune's MI block shuffle) used MathRand()%,
which is 15-bit - provably non-uniform on every full-history era over 32,768
queued samples. New 30-bit ShuffleRandomIndex().
- F2: plateau warm restarts were a no-op whenever eta already sat at its
ceiling (the normal state of a non-regressing plateau) - the ladder was just
a 24-era countdown. Restarts now overshoot to 5x the ceiling
(PLATEAU_RESTART_BOOST) and anneal geometrically back over the patience
window, SGDR-style; ETA_MIN widened 1e-4 -> 1e-5 so the decay schedule has
real range.
- F3: checkpoint restores put weights back but kept the rejected trajectory's
Adam moments, so the optimizer immediately pushed back toward the rolled-back
state (the restore->regress->restore oscillation). CNet::ResetOptimizerState()
zeroes moments/momentum/step counters (weights, BN statistics, gamma/beta
untouched) on every mid-run restore, every boosted restart, and the
deploy-time restore that online learning continues from.
- F5: batch-norm running statistics now freeze for the pass-3 OOS scoring walk,
so the selection metric the checkpoint ranking and deploy gate read is a pure
function of the checkpoint instead of partly measuring BN drift. Defensive
unfreeze in FinalizeTrainRun covers stop-mid-pass; live/online adaptation and
the OOS continual-learning simulation stay adaptive by design.
Compiled clean (0 errors, 0 warnings) via the staged-tree recipe.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The MI sample builder used `MathAbs(labelBarOffset)` as a padding, causing rows from offset and non-offset builds to be paired with a double shift. This broke the positive control, failed the 5× gate, and voided all reported mutual‑information figures. Replace with the fixed `MiShiftPad` constant to ensure builds enumerate the same set of bars and row-k alignment is preserved.
Add `BatchOptionsTotal()` to `CNeuronBatchNormOCL` and split the packed BN weight array in the learning report into separate norms for the outgoing dense matrix, gamma, beta, running statistics, and Adam moment buffers. This turns an ambiguous single‑norm reading into precise diagnostics that distinguish weight divergence from scaling issues.
- BufferDouble: replace hardcoded "DirectML/CPU-DLL" with dynamic backend name
and add buffer index/element count to all error prints for easier debugging.
- NetPersistence: distinguish missing file from transient lock by probing
FileIsExist before logging, eliminating false "sharing violation" warnings
when no saved model exists on first run.
With normalization enabled a forward pass is not a pure function of its
input - it also advances the running mean/variance. ValidateCpuInference
compares the live backend net against a throwaway pure-MQL5 clone loaded
from the just-saved .nnw, so its own reference pass left the live model one
EMA step ahead of the file the clone reads. The check would then have been
measuring its own side effect, and a marginal result decides whether
buyers' backtests are allowed to run DLL-free.
Adds CNet::SetBatchNormFrozen / CNeuronBatchNormOCL::SetStatsFrozen -
classic batch-norm inference semantics, statistics used but not updated -
and freezes both sides for the duration of the comparison. Not persisted:
it is a transient evaluation mode, not model state. Default stays
adaptive, which is what the rest of the system (online continual learning)
is built around.
Compiles 0 errors, 0 warnings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The only bounded stage in the entire forward path was the sigmoid
classification head - every hidden stage is PRELU. That is a network with
no internal scale control, and the failure ordered exactly by depth: on
SP500 H1 the shallow perceptron held ~52% balanced accuracy while the
deepest topology sat on the 33.3% one-class floor, with the per-bar logit
spread decaying monotonically (0.45 -> 0.38 over ~200 eras) until the
evidence tilt fell under the class-prior tilt. That is the signature of
internal covariate shift, which chapter 6.1 of the reference book is
entirely about and which the NeuroNet_DNG engine addresses with a layer
this project never had.
Two mechanisms make this the right fix rather than more hyperparameter
nudging:
- it decouples WEIGHT_DECAY from the learned function (van Laarhoven
2017) - with a normalized layer downstream, decay can no longer grind
the discriminative signal away, it only rescales the effective
learning rate;
- it is the precondition for ever running an unbounded logit head here.
The 2026-07-27 attempt blew up (IS error 5.6e15) precisely because
nothing upstream constrained scale.
Implementation notes:
- CNeuronBatchNormOCL computes host-side rather than as a fourth copy of
a kernel across Network.cl + WarriorCPU.cpp + WarriorDML.cpp. The math
is elementwise O(n); this way it behaves identically on all four
compute tiers, needs no DLL rebuild, and cannot drift between
backends. Same precedent as the softmax+CCE gradient and the
per-sample loss weighting, both computed in MQL5 for that reason.
- Statistics are exponential moving, not a stored mini-batch: training
is pure online SGD, one update per sample, so there is no batch to
average over. BatchNormWindow is an EMA window length.
- gamma/beta are excluded from weight decay, deliberately - decaying
gamma toward zero is the exact pathology being fixed.
- The layer self-sizes from whatever sits below it, because a conv/pool
stage's output width is derived inside the CNet constructor and is not
knowable to the topology builder.
- Checkpoint capture/restore/blend carry gamma/beta and the running
statistics alongside the dense matrix, so the plateau ladder cannot
restore a mismatched pair.
- SeedOutputLayerBias accepted only an exact defNeuronBaseOCL as the
weight-carrying penultimate layer; with normalization enabled that is
the batch-norm layer, so the cold-start bias seed would have silently
stopped being applied.
- Refuses to build, loudly, if a topology asks for normalization with no
compute backend at all - rather than quietly training a different
architecture than the one requested.
EnableBatchNorm (default on) and BatchNormWindow (1000 samples) are
inputs so the effect can be A/B'd without a recompile. Both feed the
weights-filename fingerprint, appended conditionally so existing non-BN
configs keep their fingerprints and are not forced to retrain.
Verified: analytic gradients match finite differences to 1.5e-7 relative
over 200 random cases; a faithful port of the full forward/backward chain
collapses to the 33.3% floor by era 4 without this layer and holds
36-43% with it. Compiles 0 errors, 0 warnings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
