Warrior_EA/Expert/AIBase/Inference.mqh

//+------------------------------------------------------------------+
//|                                                       Warrior_EA |
//|                                                     AnimateDread |
//|                                                                  |
//| Read-time signal production: softmax, prior calibration, class p|
//|                                                                  |
//| PARTIAL IMPLEMENTATION FILE - not standalone.                     |
//| This holds CExpertSignalAIBase method BODIES only. The class      |
//| declaration lives in Expert\ExpertSignalAIBase.mqh, which         |
//| #includes this file at the bottom, after the declaration. Do not  |
//| include it anywhere else and do not compile it on its own.        |
//|                                                                  |
//| Split out purely to make the 8216-line original navigable; the    |
//| code inside was moved verbatim, not rewritten.                    |
//+------------------------------------------------------------------+
#ifndef WARRIOR_AIBASE_INFERENCE_MQH
#define WARRIOR_AIBASE_INFERENCE_MQH
//+------------------------------------------------------------------+
//| Post-convergence "new bar" handler - see ScheduleTrainingIfNeeded()|
//| for why this exists: once m_trainingComplete is true, a plain new  |
//| bar must NOT re-enter Train()'s full era loop (which resets the    |
//| best-checkpoint/eta-decay tracking and runs real Net.backProp()    |
//| passes again, silently perturbing an already-converged model      |
//| forever, once per bar, with no way to ever actually finish). This  |
//| only refreshes the price/indicator buffers and re-runs inference   |
//| for the newest bar so dPrevSignal/the chart arrow stay current -   |
//| identical cost to what Train() does per-bar, minus every bit of    |
//| training (label caching, backProp, checkpointing).                 |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::RefreshConvergedSignal(void)
  {
   //--- Meta target: the meta head scores PROPOSED TRADES, not a bare bar window - a candidate-less
   //--- forward would also be width-mismatched against its input layer (window + descriptor). Its
   //--- live path is the S3 gate (LiveMetaGate, wired
   //--- 2026-08-19), which builds its own window + descriptor - this vote-refresh path stays closed.
   if(IsMetaTarget())
      return;
   //--- Size the buffers from what the FEATURE BUILDER actually needs, not from a date delta.
   //--- This used to be `Bars(sym, period, dtStudied, TimeCurrent()) + m_historyBars`. dtStudied is a
   //--- training watermark, and in the Strategy Tester it is loaded from a LIVE-chart save whose
   //--- timestamp is AHEAD of the simulated date - so the interval inverts, Bars() returns ~0, and the
   //--- buffer came out at exactly m_historyBars. That is just deep enough for the OHLC window to
   //--- succeed and far too shallow for the swing-context block behind it: the Donchian-50, the 20-bar
   //--- return and the SMA extension all reach further back than m_historyBars, hit the end of the
   //--- loaded series, and take their graceful degraded path. The result was silent - no error, no short
   //--- window, just inference computing DIFFERENT features from the ones training learned on. Live it
   //--- was the same bug with a milder constant (the delta is ~1 bar, giving m_historyBars + 1).
   //--- SWING_SCAN_CAP_BARS is the deepest lookback any feature performs (FindConfirmedZigZagPivot's
   //--- bound); everything else in BufferTempDataCompute reaches less far.
   int need = (int)m_historyBars + SWING_SCAN_CAP_BARS + MathMax(m_barrierHorizonBars, 1) + 2;
   int barsNow = (int)MathMin(need, Bars(m_symbol.Name(), PERIOD_CURRENT));
   //--- HOLD RATHER THAN TRADE ON A SHORT WINDOW. `need` is the depth the feature builder requires for
   //--- inference features to match the ones training learned on; a shallower buffer does not fail, it
   //--- makes the swing block take its graceful degraded path - which is precisely the silent
   //--- feature-mismatch this whole `need` calculation was introduced to end (see above). If the
   //--- indicators cannot serve `need`, the only safe output is no output: a signal computed from a
   //--- different feature distribution than the model was fitted on is worse than no signal, and this
   //--- EA sizes real positions off it.
   //---
   //--- In practice this cannot fire on a sane terminal - `need` tops out around 1,152 bars (16 + 750 +
   //--- 384 + 2) and the SMALLEST "Max bars in chart" MT5 offers is 5,000. It is insurance against the
   //--- failure mode being reachable at all, not a case expected in the field.
   int servable = ServableBars(need, "live inference");
   if(servable < need)
     {
      if(!m_inferenceDepthRefusalWarned)
        {
         m_inferenceDepthRefusalWarned = true;
         PrintFormat("%s: LIVE INFERENCE HELD - the feature window needs %d bars and the indicators can"
                     " only serve %d. Computing a signal here would silently use the swing block's"
                     " degraded path, i.e. different features from the ones this model was trained on,"
                     " so no signal is emitted until the depth is available. See the indicator-cap line"
                     " above for how to raise it.", ID, need, servable);
        }
      return;
     }
   m_inferenceDepthRefusalWarned = false;
   if(!ResizeBuffers(barsNow) || !RefreshData())
      return;
   //--- INVALIDATE THE NOW-RELATIVE BAR CACHES. Non-obvious and load-bearing: the feature cache is keyed
   //--- by MQL5 series index, and index 0 means "newest bar", so every closed candle shifts what every
   //--- cached row stands for. Train() is the only other caller of this, and once m_trainingComplete is
   //--- set ScheduleTrainingIfNeeded() routes every subsequent bar HERE instead - Train() is never
   //--- re-entered, so without this call nothing ever clears the cache again for the rest of the process.
   //--- A chart that trained to convergence (or was deployed via DeployNow()) would then keep replaying
   //--- the rows computed for the last training era's bar grid: BufferTempData(0..m_historyBars-1) all hit
   //--- the cache, the feature window never changes, and dPrevSignal freezes at its convergence-time value
   //--- forever - silently, since every buffer above refreshed correctly and the vector is the right SHAPE.
   //--- OnlineLearnStep() below would compound it by backpropping those stale features against freshly
   //--- resolved labels, i.e. actively training the deployed model on mismatched pairs.
   //--- A freshly started inference-only process (a backtest, or a buyer loading a deployed .nnw) was
   //--- never affected: it never allocates these arrays, so BufferTempData()'s `cacheable` test is false
   //--- and it always recomputes. This is a live/forward-chart fix, not a backtest one.
   EnsureBarCachesCapacity(barsNow);
   //--- Same bar grid, same panel. A deployed model never enters Train(), so this is the only place
   //--- its cross-asset panel gets built - and it must be built from the SAME reference set training
   //--- used, or inference reads a different feature vector than the weights were fitted to.
   //--- Only as deep as inference actually reads. RefreshLatestSignal() touches bars 1..m_historyBars
   //--- (window ends on the newest CLOSED bar - the +2 slack below covers the extra bar of depth)
   //--- and the panel's own slow window reaches CROSSASSET_SLOW_BARS further back - nothing else. Asking
   //--- for the full `barsNow` here would rebuild a training-depth panel on EVERY bar, which in the
   //--- tester means one full multi-symbol resample per simulated bar. The cache check in
   //--- BuildCrossAssetPanel is >=, so a deeper panel left over from training still satisfies this.
   BuildCrossAssetPanel((int)m_historyBars + CROSSASSET_SLOW_BARS + 2);
   EnsureSpreadSeries(barsNow);
   //--- A deployed model never enters Train(), so this is the only place its barrier horizon gets
   //--- measured - and OnlineLearnStep() below depends on it being right. First call sizes buffers
   //--- against the fallback, which is harmless: `need` is dominated by SWING_SCAN_CAP_BARS either way.
   EnsureBarrierHorizon(barsNow);
   bool refreshed = RefreshLatestSignal();
   //--- Continual learning: on a LIVE chart (never the tester/optimizer - OnlineLearnStep() self-guards
   //--- on m_inferenceOnly) a deployed model keeps adapting to newly-confirmed structure. Runs AFTER the
   //--- live signal is drawn (so the arrow uses the shadow as it was for THIS bar's decision) and BEFORE
   //--- dtStudied advances (OnlineLearnStep keeps its own time watermark, independent of dtStudied).
   OnlineLearnStep();
   //--- Advance the live new-bar watermark ONLY on success. Advancing it unconditionally meant a
   //--- transient window failure (indicator hole, history hiccup) closed the gate for the rest of the
   //--- bar with the PREVIOUS bar's dPrevSignal still voting - the tester path (m_lastBarTime) already
   //--- advanced only on success and self-healed; this is the live path catching up. On failure the
   //--- gate stays open, so the next tick retries.
   if(refreshed)
      dtStudied = m_Time.GetData(0);
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
bool CExpertSignalAIBase::RefreshLatestSignal(void)
  {
   //--- Meta target: the live path is the S3 gate (LiveMetaGate), not the per-bar vote - see
   //--- RefreshConvergedSignal's meta guard.
   if(IsMetaTarget())
      return false;
   //--- Bar 1: the newest CLOSED bar, NOT the forming bar. This runs at the first tick after a bar
   //--- opens, when series index 0 is a bar with one tick of data: (close-open)/atr ~ 0, high ~ low,
   //--- a degenerate volume block, indicators computed on a 1-tick candle. Train() never produces
   //--- such a window - every labeled bar is fully closed, and its label assumes entry at that bar's
   //--- CLOSE (see TripleBarrierLabel's header). The training-parity query at this instant (fixed
   //--- 2026-08-11) is therefore the window ending on bar 1, whose close IS the current price - the
   //--- exact instant the label's hypothetical entry happens. The old i = 0 fed the deployed model an
   //--- out-of-distribution final timestep - the timestep the LSTM/HYBRID output is keyed to - and
   //--- semantically asked for the label of a bar whose close was still an hour away, so the deploy
   //--- gate's OOS scores (closed bars, pass 3) measured a different query than live executed. Both
   //--- paths go through BuildFeatureWindow(), which guarantees identical construction; this index is
   //--- what makes them the same QUESTION.
   int i = 1;
   int r = i;
   if(!BuildFeatureWindow(r))
     {
      //--- One combined failure now (partial window OR short total) where there used to be two counters.
      //--- Kept distinct in the tally by testing what actually landed: a window that built every bar but
      //--- came up short is the "short" case, anything else is a feature-build failure.
      if(TempData.Total() > 0 && TempData.Total() < (int)m_historyBars * m_neuronsCount)
         m_refreshFailShort++;
      else
         m_refreshFailFeatures++; // see PrintInferenceTally()
      //--- No opinion this bar rather than a stale one: dPrevSignal still holds the PREVIOUS bar's
      //--- decision, and LongCondition()/ShortCondition() would keep voting that stale direction all
      //--- bar. The caller retries (RefreshConvergedSignal only advances dtStudied on success), so a
      //--- transient failure costs ticks, not the bar.
      dPrevSignal = 0.0;
      return false;
     }
   //--- Live trading/inference reads from the EMA shadow net, not Net directly - see m_shadowNet's
   //--- declaration comment. Falls back to Net if the shadow isn't bootstrapped yet (should only be
   //--- momentarily, on a genuinely fresh start before EnsureShadowNet() has run).
   EnsureShadowNet();
   CNet *deployNet = (CheckPointer(m_shadowNet) != POINTER_INVALID) ? m_shadowNet : Net;
   deployNet.feedForward(TempData);
   deployNet.getResults(TempData);
   if(m_outputNeuronsCount == 1)
      dPrevSignal = TempData[0];
   else
      if(m_outputNeuronsCount == 3)
        {
         //--- Live decision. ApplyClassificationSoftmax() computes the softmax INTO TempData and returns
         //--- the decision; AdjustedSignalFromSoftmax() re-reads that same TempData and applies the same
         //--- strict-majority/ties-to-Neutral rule, so since the read-time prior correction was removed
         //--- (2026-07-31) the two provably agree. The call is kept because a dozen sites name it as
         //--- "the live decision rule" and that is still exactly what it is - the correction now lives
         //--- in the trained weights instead of here.
         //--- The "raw softmax was neutralized by prior correction" diagnostic that used to sit here went
         //--- with it: with nothing between the two values it could never fire again.
         ApplyClassificationSoftmax();
         dPrevSignal = AdjustedSignalFromSoftmax();
        }
   m_refreshOk++;
   //--- bt anchors the DECISION bar (bar 1, the closed bar the window ends on) - it keys the arrow,
   //--- its High/Low placement and NMS declustering, and now matches the rescan path, which draws
   //--- each historical arrow at the bar its window ends on.
   datetime bt = m_Time.GetData(i);
   //--- Keep a pure inference-side watermark of the newest bar FRAME this model has already evaluated.
   //--- This must be the FORMING bar's open time (index 0), not bt: the new-bar gate compares it
   //--- against SERIES_LASTBAR_DATE (also the forming bar's open), so anchoring it at bt (bar 1)
   //--- would compare one bar behind and re-fire the refresh on every tick forever. The tester may
   //--- load dtStudied from a live-chart save whose timestamp is AHEAD of the simulated backtest date
   //--- range; using that training watermark to decide whether a "new bar" exists then freezes
   //--- dPrevSignal at its init-bar value for the whole run. m_lastBarTime is this runtime's own
   //--- latest evaluated frame instead, so it stays aligned to whichever history the current process
   //--- is actually traversing.
   m_lastBarTime = m_Time.GetData(0);
   //--- LIVE NMS, AND IT NOW GATES THE TRADE, NOT JUST THE ARROW.
   //---
   //--- It used to sit at the bottom of this function wrapped around DrawObject() alone, so a suppressed
   //--- bar lost its arrow and still traded: dPrevSignal was never touched, and dPrevSignal is what
   //--- LongCondition()/ShortCondition()/SignedAIConfidence() read. The chart therefore showed roughly one
   //--- arrow per EIGHT positions the EA would open - measured on SP500 H1 2026-08-09, where CONV called a
   //--- direction on 64% of bars while ~40 arrows appeared across the ~500 visible ones. Worse, the arrows
   //--- that survived were not a random eighth: rule 2 below keeps the HIGHER-CONFIDENCE side of a
   //--- cluster, so the visible set was systematically the best member of each run. A chart that shows the
   //--- best of every eight decisions and hides the rest reads far better than the model is, which is the
   //--- same best-of-N selection error this codebase has now corrected in four other places - this time on
   //--- the display layer, where it is most likely to mislead the person deciding whether to trade it.
   //---
   //--- Neutralising dPrevSignal (rather than adding a separate "may trade" flag consulted at each of the
   //--- half-dozen read sites) is deliberate: it leaves exactly ONE definition of what this model decided
   //--- this bar, so the arrow, the panel's "Current signal", the confidence handed to sizing/SL/TP/
   //--- trailing, the refresh tally below and the order itself cannot drift apart again. One arrow is now
   //--- one trade, which is what makes the chart an honest record.
   //---
   //--- NOTE the scoring consequence, deliberately NOT papered over: the era line's dir-precision still
   //--- counts EVERY directional call, so it now describes a larger population than the one that trades.
   //--- The era line carries a separate declustered figure alongside it (see m_oosNmsFired) so both are
   //--- visible; the selection metric is not switched over until those numbers show what the coverage
   //--- floor should be, because a blind switch is how the minRR and recall-floor catch-22s happened.
   ENUM_SIGNAL lsig = DoubleToSignal(dPrevSignal);
   bool nmsAccept = (lsig != Neutral) && NmsLiveAccept(bt, lsig, MathAbs(dPrevSignal));
   if(lsig != Neutral && !nmsAccept)
      dPrevSignal = 0.0;   // declustered away: no arrow, no vote, no position
   switch(DoubleToSignal(dPrevSignal))
     {
      case Buy:
         m_refreshBuy++;
         break;
      case Sell:
         m_refreshSell++;
         break;
      default:
         m_refreshNeutral++;
         break;
     }
   if(nmsAccept)
      DrawObject(bt, dPrevSignal, m_Close.GetData(i));
   else
      DeleteObject(bt);
   return true;
  }
//+------------------------------------------------------------------+
//|                                                                  |
//+------------------------------------------------------------------+
double CExpertSignalAIBase::ApplyClassificationSoftmax(void)
  {
   // A non-finite logit poisons everything downstream: maxLogit, every exp(), the sum, and all three
   // probabilities become NaN, and since NaN fails every comparison the two directional tests below
   // are both false - so a NaN'd net returns Neutral on every bar forever and looks EXACTLY like a
   // model that has simply gone quiet. That is the failure mode this project has chased repeatedly
   // from the outside (panel says "no directional calls", nobody can tell whether the model is
   // cautious or dead). Detect it here, at the one place the raw logits are first read, and say so.
   if(!MathIsValidNumber(TempData.At(0)) || !MathIsValidNumber(TempData.At(1)) || !MathIsValidNumber(TempData.At(2)))
     {
      static int nanLogitReports = 0;
      // Bounded: this cannot heal on its own (the weights are already corrupt), so unlimited logging
      // would fill the journal for as long as the chart stays attached. Three is enough to prove it.
      if(nanLogitReports < 3)
        {
         nanLogitReports++;
         PrintFormat("%s: %s NON-FINITE network output (%g / %g / %g) - forcing Neutral. The weights are "
                     "corrupt; reload the last good .nnw or reset and retrain. Report %d of 3.",
                     __FUNCTION__, ID, TempData.At(0), TempData.At(1), TempData.At(2), nanLogitReports);
        }
      return 0;
     }
   // CLASS_LOGIT_SCALE (AI\Network.mqh) must match the training-gradient softmax in
   // backProp/backPropOCL exactly - this is the same normalization the loss was trained against.
   double maxLogit = CLASS_LOGIT_SCALE * MathMax(TempData.At(0), MathMax(TempData.At(1), TempData.At(2)));
   double sum = 0;
   for(int res = 0; res < 3; res++)
     {
      double temp = exp(CLASS_LOGIT_SCALE * TempData.At(res) - maxLogit);
      sum += temp;
      TempData.Update(res, temp);
     }
   for(int res = 0; res < 3; res++)
      TempData.Update(res, TempData.At(res) / sum);
   double pBuy = TempData.At(0);
   double pSell = TempData.At(1);
   double pNeutral = TempData.At(2);
   // TempData.Maximum(0,3) scans left-to-right and keeps the FIRST index on a tie, so any tie
   // (including the degenerate all-equal 0.3333/0.3333/0.3333 case from a collapsed/untrained net)
   // always resolved to Buy (index 0) - silently turning "the model has no idea" into a directional
   // trade. Buy/Sell now only win with a strict majority over BOTH other classes; every tie,
   // 2-way or 3-way, falls through to Neutral.
   if(pBuy > pSell && pBuy > pNeutral)
      return pBuy; // Buy signal
   if(pSell > pBuy && pSell > pNeutral)
      return -pSell; // Sell signal
   return 0; // Neutral signal (also the fallback on any tie)
  }
//+------------------------------------------------------------------+
//| Post-hoc logit adjustment (prior correction) of the 3-class      |
//| decision. Reads the raw softmax probabilities ApplyClassification|
//| Softmax() left in TempData[0..2] and returns the prior-corrected |
//| signed decision (+P'(buy)/-P'(sell)/0-neutral), the exact rule    |
//| live trading fires on and the live-fired precision metric scores. |
//|                                                                   |
//| RAW ARGMAX, deliberately. The prior correction this function used |
//| to apply at read time (Saerens et al. 2002) was REMOVED           |
//| 2026-07-31 along with the AILogitPriorStrength input.             |
//|                                                                   |
//| Why there is nothing to correct here: the logit-adjusted loss     |
//| adds tau*log(prior_c) to each class logit inside the TRAINING     |
//| gradient, so the network learns to absorb the offset and its raw  |
//| argmax is ALREADY the balanced-error-optimal decision. Applying a |
//| second correction at inference would account for the same base    |
//| rate twice and push the decision back toward Neutral - undoing    |
//| exactly what the loss bought. The old code knew this: the whole   |
//| adjustment sat behind an `if(m_useLogitAdjustedLoss) return raw`  |
//| guard and had been unreachable for the entire shipped default     |
//| configuration. Kept as a named function rather than inlined       |
//| because a dozen call sites document themselves by calling "the    |
//| live decision rule" - and that is exactly what this is.           |
//+------------------------------------------------------------------+
double CExpertSignalAIBase::AdjustedSignalFromSoftmax(void)
  {
   if(TempData.Total() < 3)
      return 0.0;
   double pBuy = TempData.At(0), pSell = TempData.At(1), pNeutral = TempData.At(2);
   //--- Strict majority, ties to Neutral - the same rule as ApplyClassificationSoftmax(). The returned
   //--- magnitude is a genuine probability, which the confidence floor and ConfidenceTier() read.
   bool wantBuy  = (pBuy > pSell && pBuy > pNeutral);
   bool wantSell = (pSell > pBuy && pSell > pNeutral);
   if(!wantBuy && !wantSell)
      return 0.0;
   //--- OPERATING POINT (2026-08-09). Argmax alone answers "which class is most likely"; it does not
   //--- answer "is this worth trading", and those are different questions whenever the top two classes
   //--- are nearly tied. A marginal directional win over Neutral used to become a trade, which is the
   //--- mechanical source of the model calling a direction on ~90% of bars. Below the fitted margin
   //--- this abstains instead - and abstaining is not a loss of information, it is the model declining
   //--- to act on a distinction it cannot make. See DIR_CONF_THRESHOLD_BINS for how the value is chosen.
   //---
   //--- Returning Neutral rather than exposing a separate "tradeable" flag is deliberate, and matches
   //--- the same decision made for live NMS (see RefreshLatestSignal): one definition of what this model
   //--- decided this bar, so the arrow, the panel, the confidence handed to sizing/SL/TP, the OOS score
   //--- and the order itself cannot drift apart.
   if(m_dirConfThreshold > 0.0)
     {
      double win    = wantBuy ? pBuy : pSell;
      double rival  = wantBuy ? MathMax(pSell, pNeutral) : MathMax(pBuy, pNeutral);
      if((win - rival) < m_dirConfThreshold)
         return 0.0;
     }
   return wantBuy ? pBuy : -pSell;
  }
//+------------------------------------------------------------------+
//| The statistic the operating point is expressed in - see the       |
//| declaration. Reads the softmax ALREADY in TempData, so callers    |
//| must have run ApplyClassificationSoftmax() first.                 |
//+------------------------------------------------------------------+
double CExpertSignalAIBase::DirectionalMargin(void)
  {
   if(TempData.Total() < 3)
      return -1.0;
   double pBuy = TempData.At(0), pSell = TempData.At(1), pNeutral = TempData.At(2);
   if(pBuy > pSell && pBuy > pNeutral)
      return pBuy - MathMax(pSell, pNeutral);
   if(pSell > pBuy && pSell > pNeutral)
      return pSell - MathMax(pBuy, pNeutral);
   return -1.0;   // Neutral won: no directional call, so no operating point applies
  }
//+------------------------------------------------------------------+
//| Clear the margin histogram at the start of the calibration walk.  |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::ResetDirConfHistogram(void)
  {
   ArrayInitialize(m_dirConfBinCalls, 0);
   ArrayInitialize(m_dirConfBinHits, 0);
   m_dirConfPrimaryBars = 0;
  }
//+------------------------------------------------------------------+
//| One calibration sample. isPrimaryBar survives from when this was  |
//| harvested inside pass 2's oversampled replay queue, where counting |
//| duplicated minority bars would have fitted the operating point to  |
//| a class balance the live model never sees (the same correction     |
//| m_cumIsTotal makes - see its note in Training.mqh). The calibration |
//| walk visits each bar exactly once and passes true; the parameter   |
//| stays so any future caller must state which it is.                 |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::AccumulateDirConfSample(double margin, bool wasCorrect, bool isPrimaryBar)
  {
   if(!isPrimaryBar)
      return;
   //--- Counted BEFORE the directional test: this is the coverage denominator, so it has to be every
   //--- primary bar the model scored, including the ones it called Neutral. Using only directional
   //--- bars would make coverage 100% by construction at every threshold.
   m_dirConfPrimaryBars++;
   if(margin < 0.0)
      return;                                  // Neutral won - not a directional call
   int bin = (int)(margin * DIR_CONF_THRESHOLD_BINS);
   if(bin < 0)
      bin = 0;
   if(bin >= DIR_CONF_THRESHOLD_BINS)
      bin = DIR_CONF_THRESHOLD_BINS - 1;       // margin can reach exactly 1.0
   m_dirConfBinCalls[bin]++;
   if(wasCorrect)
      m_dirConfBinHits[bin]++;
  }
//+------------------------------------------------------------------+
//| Choose the operating point: the margin that maximises EXPECTANCY  |
//| on the held-out calibration slice while still calling a direction |
//| often enough to clear the SAME coverage floor the deploy gate     |
//| uses. Held-out matters as much as the objective does - see        |
//| DIR_CONF_CALIB_PCT_OF_IS for what fitting it on the training      |
//| bars did to the sign of (p - break-even).                         |
//|                                                                   |
//| Swept from the top down so the running totals are "calls at or    |
//| above this bin", which is exactly the set a threshold there would |
//| admit - one pass, no nested loop over candidate thresholds.       |
//|                                                                   |
//| TIES GO TO THE LOWER THRESHOLD. Precision is a ratio of counts    |
//| and plateaus over ranges of margin; taking the highest threshold  |
//| on a plateau would buy identical precision for strictly less      |
//| coverage, and coverage is what keeps the model tradeable.         |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::FitDirConfThreshold(void)
  {
   long totalCalls = 0;
   for(int b = 0; b < DIR_CONF_THRESHOLD_BINS; b++)
      totalCalls += m_dirConfBinCalls[b];
   if(totalCalls < DIR_CONF_MIN_FIT_CALLS || m_dirConfPrimaryBars <= 0)
     {
      //--- Not enough evidence to place an operating point. KEEP THE PREVIOUS ONE - the old behaviour
      //--- here was to reset to 0.0, which is "call a direction on every bar", the single most exposed
      //--- setting in the range. A failed measurement must never decay to the most aggressive value it
      //--- could have returned; the last threshold that WAS fitted is a strictly better estimate than
      //--- the one setting we know maximises exposure. At era 0 the previous value is 0.0 regardless,
      //--- so the cold-start path is unchanged.
      if(!m_dirConfSparseWarned)
        {
         m_dirConfSparseWarned = true;
         Print(ID + StringFormat(": directional confidence threshold NOT refitted - only %d directional "
                                 "calls in the held-out calibration slice this era (need %d). Keeping "
                                 "the previous operating point %.2f; this is normal for the first eras "
                                 "and self-corrects as the model starts calling directions.",
                                 (int)totalCalls, DIR_CONF_MIN_FIT_CALLS, m_dirConfThreshold));
        }
      return;
     }
   //--- The floor is the true directional base rate x MIN_COVERAGE_FRACTION_OF_BASE_RATE, matching
   //--- Train()'s minCoveragePct exactly. Derived from THIS era's own IS labels rather than passed in,
   //--- so the two cannot fall out of step when one of them is edited.
   long trueDir = m_trueBuyCount + m_trueSellCount;
   long trueTot = trueDir + m_trueNeutralCount;
   double baseRatePct = (trueTot > 0) ? 100.0 * (double)trueDir / trueTot : 0.0;
   double minCoveragePct = baseRatePct * MIN_COVERAGE_FRACTION_OF_BASE_RATE;
   //--- EXPECTANCY, NOT PRECISION. Maximising the win rate alone has no answer for a PLATEAU, and the
   //--- previous `precPct >= bestPrec` resolved one by walking to ever more coverage. That is a
   //--- catastrophe on exactly the models that need a threshold most: a net with no edge scores its base
   //--- rate at EVERY threshold, which is a perfect plateau, so the walk ran to bin 0 and returned
   //--- threshold 0.0 - fire on every bar. Observed 2026-08-10 as PAI "overshooting signals" while the
   //--- other three stayed selective; PAI has the most degenerate margin distribution (OOS outputs
   //--- spanning the full 0.000..1.000 where CONV sits at 0.214..0.814), so its plateau is the flattest.
   //---
   //--- The money quantity is expectancy per BAR, and for a k:m barrier
   //---   EV = (p - p0) * (k + m)  with p0 = m/(m+k),
   //--- so EV per bar = coverage * (p - p0) * (k + m). (k+m) is constant across thresholds, which
   //--- leaves coverage * (p - p0) as the objective. It behaves correctly in all three regimes and
   //--- needs no tie-break rule:
   //---   p > p0 everywhere -> more coverage is more money, so it takes the coverage (the old
   //---                        behaviour, but for a reason rather than as a plateau artifact)
   //---   p flat AT p0      -> every point scores 0 and the floor decides; no runaway
   //---   p < p0 everywhere -> the LEAST coverage loses the least, so it becomes MORE selective
   //---                        instead of trading everything, which is the current reality for all
   //---                        four models and the opposite of what the old rule did.
   //--- COST-ADJUSTED, 2026-08-17. This is the reference the operating-point objective subtracts, so
   //--- using the frictionless SL/(SL+TP) here made every candidate threshold look better than it was by
   //--- the width of the spread - on SP500 H4 that was 2.2pp against a measured edge of 2.3pp, i.e. very
   //--- nearly all of it. See CostAdjustedBreakEvenPct.
   double breakEvenPct = CostAdjustedBreakEvenPct();
   //--- Per-bin curve, cached so the second pass does not re-accumulate. Values are the running
   //--- "at or above this bin" totals, which is exactly the population a threshold there admits.
   double binCov[DIR_CONF_THRESHOLD_BINS];
   double binPrec[DIR_CONF_THRESHOLD_BINS];
   double binScore[DIR_CONF_THRESHOLD_BINS];
   double binSe[DIR_CONF_THRESHOLD_BINS];
   bool   binOk[DIR_CONF_THRESHOLD_BINS];
   long   runCalls = 0, runHits = 0;
   double bestScore = -DBL_MAX, bestSe = 0.0;
   int    bestBin = -1, floorBin = -1, eligibleBins = 0;
   for(int b = DIR_CONF_THRESHOLD_BINS - 1; b >= 0; b--)
     {
      binOk[b] = false;
      binCov[b] = 0.0;
      binPrec[b] = 0.0;
      binScore[b] = 0.0;
      binSe[b] = 0.0;
      runCalls += m_dirConfBinCalls[b];
      runHits  += m_dirConfBinHits[b];
      if(runCalls <= 0)
         continue;
      double coveragePct = 100.0 * (double)runCalls / m_dirConfPrimaryBars;
      if(coveragePct < minCoveragePct)
         continue;                              // too selective to be deployable
      double precPct = 100.0 * (double)runHits / runCalls;
      double p = precPct / 100.0;
      binOk[b] = true;
      binCov[b] = coveragePct;
      binPrec[b] = precPct;
      binScore[b] = coveragePct * (precPct - breakEvenPct);
      //--- Binomial standard error of the win rate at this operating point, carried into the score's
      //--- own units. Coverage is measured against a FIXED denominator every bin, so it is far better
      //--- determined than the win rate; the score's error is dominated by the precision term.
      //--- ON THE EFFECTIVE SAMPLE, not the raw call count (2026-08-17). These calls are triple-barrier
      //--- outcomes on consecutive bars, so they overlap: at a 384-bar horizon, neighbouring labels share
      //--- almost their entire outcome window and are nothing like independent draws, and runCalls
      //--- understates the error by up to ~sqrt(mean lifespan).
      //--- NOT because the gate below was observed to misfire - measured over the full 6,930-era run it
      //--- fires on 1.5-8.2% of eras, at or under the ~5% a family-wise test should. See
      //--- EffectiveSampleSize(); this is a formula correction, not a bug fix, and it makes the bar
      //--- higher rather than lower.
      binSe[b] = coveragePct * 100.0 * MathSqrt(MathMax(p * (1.0 - p), 0.0)
                                                / EffectiveSampleSize((double)runCalls));
      //--- The sweep runs top-down, so the FIRST eligible bin is the most selective one that still
      //--- clears the coverage floor. That point is the deterministic fallback below.
      if(floorBin < 0)
         floorBin = b;
      eligibleBins++;
      //--- Strict >, so a genuine tie keeps the MORE selective point (the loop reaches it first). The
      //--- old >= did the reverse and that is what made the plateau run away.
      if(binScore[b] > bestScore)
        {
         bestScore = binScore[b];
         bestSe = binSe[b];
         bestBin = b;
        }
     }
   if(bestBin < 0)
     {
      //--- Even calling on every directional argmax does not reach the coverage floor, so there is no
      //--- room to be MORE selective. Unthresholded is then the only setting that can clear the gate.
      m_dirConfThreshold = 0.0;
      return;
     }
   //--- SELECTION UNDER A NULL OF THE MAXIMUM, with a parsimony fallback in the spirit of the
   //--- one-standard-error rule (Breiman et al. 1984, CART 3.4.3; Hastie/Tibshirani/Friedman, ESL 2ed
   //--- 7.10 - prefer the simpler model when the score difference is inside the noise). The bare
   //--- argmax above is the right ANSWER only if the curve it maximises is measured well enough to
   //--- rank its own candidates, and on this data it is not. Measured over 98 consecutive fits of the
   //--- shipped SP500 H4 model:
   //---
   //---   correlation(chosen threshold, win rate at it) = -0.056 over the full 0.00..0.74 range
   //---   win rate stdev across fits                    = 1.32pp
   //---   binomial SE of that win rate at ~1430 calls    = 1.25pp
   //---
   //--- The correlation is zero - the margin does not rank trades at all - and the era-to-era spread
   //--- IS its own sampling error to within 0.07pp. So `coverage x (precision - breakEven)` was
   //--- `coverage x (3.4 +/- 1.3)`, and taking the argmax over ~37 eligible bins returned whichever
   //--- bin drew the luckiest sample. The threshold then teleported 0.42 -> 0.04 -> 0.74 in three
   //--- eras, swinging OOS coverage 0% -> 39%, which left the era win rate measured on 1-5 calls and
   //--- swinging 0% <-> 100%. That is the whole of the "training is highly unstable" report, and none
   //--- of it was the optimizer.
   //---
   //--- This is the same defect the family-wise gate rule already governs elsewhere in this file - a
   //--- best-of-N adopted without a null of the maximum - applied here to the operating point rather
   //--- than the deploy decision.
   //---
   //--- THE RULE. The argmax is adopted only if it beats the DETERMINISTIC fallback by more than a
   //--- best-of-N maximum could manage on noise alone; otherwise the fallback is taken.
   //---
   //--- Fallback = the most selective bin that still clears the coverage floor. That point is chosen
   //--- from the MARGIN DISTRIBUTION only - it never consults a win rate - so it carries none of the
   //--- outcome noise that was driving the thrash, and it moves era to era only when the model's own
   //--- confidence distribution genuinely moves. It is also the conservative end of the sweep: the
   //--- fewest bars called that still leaves a deployable model, which is the right default on a
   //--- funded account when no operating point has been shown to be better than another.
   //---
   //--- A plain one-standard-error band was the first thing tried here and it is NOT sufficient: the
   //--- band edge is bestScore - bestSE, and with an edge of 2.3pp against a 1.25pp standard error
   //--- bestScore is itself +/-50%, so the admitted set - and the coverage it implies - would still
   //--- wander by half its own width every era. The fallback has to be independent of the noisy
   //--- quantity, not merely a wider window around it.
   //---
   //--- Significance uses the null of the MAXIMUM, not a per-candidate test: the argmax is the best of
   //--- `eligibleBins` draws, and the expected maximum of N standard normals grows like sqrt(2 ln N),
   //--- so that is the bar it has to clear. Same correction the deploy gate already applies to
   //--- best-of-N model selection, applied here to the operating point.
   double refScore = binScore[floorBin];
   double refSe = binSe[floorBin];
   //--- Conservative: the two points are NESTED samples, so their difference is better determined than
   //--- this independent-errors sum implies. Erring toward "not significant" is the safe direction.
   double seDiff = MathSqrt(bestSe * bestSe + refSe * refSe);
   double zMax = MathSqrt(2.0 * MathLog(MathMax((double)eligibleBins, 2.0)));
   bool   separates = ((bestScore - refScore) > zMax * seDiff);
   int    selBin = (separates ? bestBin : floorBin);
   double bestThresh = (double)selBin / DIR_CONF_THRESHOLD_BINS;
   double bestPrec = binPrec[selBin];
   double bestCov = binCov[selBin];
   double prevThresh = m_dirConfThreshold;
   m_dirConfThreshold = bestThresh;
//--- Built as a local rather than inlined into the ternary: the two branches are long enough that
//--- keeping them out of the argument list is what makes the call readable.
   string ruleNote = "which CLEARS that bar, so the margin genuinely separates these operating points"
                     " and the argmax was adopted";
   if(!separates)
      ruleNote = "which it does NOT clear - the margin does not rank these trades, so the operating"
                 " point fell back to the most selective bin that still clears the coverage floor."
                 " That fallback reads the margin DISTRIBUTION only, never a win rate, so it cannot"
                 " thrash on outcome noise the way the argmax did";
//--- Logged when it moves a bin AND the era cadence is due (2026-08-19). "Stays quiet when stable"
//--- had stopped being a filter: the operating point is measured noise-dominated (project memory:
//--- "ratchet, then noise"), so it moved a bin nearly every era - ~400 prints per member per day.
//--- The threshold itself keeps updating every era regardless; only the announcement is throttled.
   if(TrainLogDue() && MathAbs(m_dirConfThreshold - prevThresh) >= 1.0 / DIR_CONF_THRESHOLD_BINS)
      Print(ID + StringFormat(": directional confidence threshold %.2f -> %.2f (fitted on %d HELD-OUT "
                              "calibration bars: %.1f%% coverage at %.1f%% WIN RATE vs " + DoubleToString(breakEvenPct, 1) +
                              "%% break-even, edge " + DoubleToString(bestPrec - breakEvenPct, 1) +
                              "pp, coverage floor %.1f%%). Below "
                              "this winner-vs-rival margin the model abstains instead of trading. The "
                              "rate is wins - target before stop on the side actually called - not "
                              "agreement with the collapsed 3-class label; see m_oosBuyPredictedWins."
                              " | best-of-N gate over %d eligible bins: argmax %.2f scored %.0f vs the"
                              " %.2f fallback's %.0f, a gap of %.0f against a null-of-the-maximum bar"
                              " of %.0f (z_max %.2f x SE %.0f, sized on the EFFECTIVE sample - labels"
                              " overlap by a mean lifespan of %.0f bars, so n is deflated %.0fx), %s",
                              prevThresh, m_dirConfThreshold, (int)m_dirConfPrimaryBars, bestCov,
                              bestPrec, minCoveragePct, eligibleBins,
                              (double)bestBin / DIR_CONF_THRESHOLD_BINS, bestScore,
                              (double)floorBin / DIR_CONF_THRESHOLD_BINS, refScore,
                              bestScore - refScore, zMax * seDiff, zMax, seDiff,
                              MeanLabelLifespan(), MeanLabelLifespan(), ruleNote));
  }
//+------------------------------------------------------------------+
//| EMA-updates the persisted true class base rates from a finished   |
//| era's true class counts. First real measurement seeds directly;   |
//| thereafter blended with the same smoothing as the accuracy/       |
//| confidence EMAs so one noisy era can't swing the live decision.   |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::UpdateClassPriors(long buyCnt, long sellCnt, long neutralCnt)
  {
   long tot = buyCnt + sellCnt + neutralCnt;
   if(tot <= 0)
      return;
   double pb = (double)buyCnt / tot, ps = (double)sellCnt / tot, pn = (double)neutralCnt / tot;
   if(m_priorNeutral <= 0.0) // first real measurement
     {
      m_priorBuy = pb;
      m_priorSell = ps;
      m_priorNeutral = pn;
      return;
     }
   //--- (Was `m_useStaticPrior || m_freezePriorCalibration`. Those were two separate user-facing inputs
   //--- whose only effect anywhere in the codebase was this one OR - two controls for one decision.
   //--- UseStaticPrior was removed 2026-07-31; see the class-imbalance audit in Variables\Inputs.mqh.)
   if(m_freezePriorCalibration)
      return;
   double k = Net.recentAverageSmoothingFactor;
   if(k < 1.0)
      k = 1.0;
   m_priorBuy     += (pb - m_priorBuy) / k;
   m_priorSell    += (ps - m_priorSell) / k;
   m_priorNeutral += (pn - m_priorNeutral) / k;
  }
//+------------------------------------------------------------------+
//| Installs the training-time logit offsets - see the declaration.   |
//+------------------------------------------------------------------+
void CExpertSignalAIBase::ApplyLogitAdjustment(void)
  {
   if(CheckPointer(Net) == POINTER_INVALID)
      return;
   if(m_logitAdjustTau <= 0.0)
     {
      //--- Clear rather than merely skip: the input can be turned off on a chart that already installed
      //--- offsets this session, and a stale adjustment would keep biasing the gradient silently.
      Net.ClearLogitAdjustment();
      return;
     }
   //--- Priors not measured yet (era 0 before the first tally, or a model with no .stats): leave the
   //--- gradient unadjusted rather than guessing a distribution. The next era installs them.
   //--- THIS USED TO BE SILENT, and that silence hid a whole-run failure: while the auto-tune search ran,
   //--- UpdateClassPriors() was skipped in eval mode, so this branch was taken on EVERY era and the
   //--- imbalance correction never once ran - with nothing in the log to say so. A mechanism that
   //--- declines to act must announce it; the alternative is indistinguishable from working. Third time
   //--- this codebase has been bitten by a quiet no-op, so it now warns every time it is not merely the
   //--- expected era-0 case.
   if(m_priorBuy <= 0.0 || m_priorSell <= 0.0 || m_priorNeutral <= 0.0)
     {
      if(m_eraCount > 0 && !m_logitAdjustSkipWarned)
        {
         m_logitAdjustSkipWarned = true;
         Print(ID + ": WARNING - class-imbalance correction is NOT running at era " +
               IntegerToString(m_eraCount) + ": the class priors have never been measured (Buy " +
               DoubleToString(m_priorBuy, 4) + " Sell " + DoubleToString(m_priorSell, 4) + " Neutral " +
               DoubleToString(m_priorNeutral, 4) + "). Training is falling back to plain cross-entropy, "
               "which on a skewed label set collapses to the majority class.");
        }
      Net.ClearLogitAdjustment();
      return;
     }
   //--- Effective tau, capped so the offsets cannot swamp the head's usable logit range - see
   //--- LOGIT_ADJUST_MAX_RANGE_FRACTION. The binding quantity is the SPREAD between the largest and
   //--- smallest offset, not their absolute size: softmax is shift-invariant, so a constant added to
   //--- all three classes changes nothing and only their differences move the decision.
   double lb = MathLog(m_priorBuy), ls = MathLog(m_priorSell), lnn = MathLog(m_priorNeutral);
   //--- THE CORRECTION SPANS THE DECIDABLE CLASSES ONLY - Buy against Sell. Neutral is excluded, and
   //--- that exclusion is the whole point of this block (2026-08-16).
   //---
   //--- Logit adjustment (Menon et al. 2020) makes the classifier Bayes-optimal for BALANCED error by
   //--- subsidising rare classes. It was wired here when Neutral was the DOMINANT class - the era of
   //--- "big move up / big move down / nothing much", where the majority outcome was no move and the
   //--- correction pulled the model off it. The triple-barrier relabel (b4a704d) inverted that: the
   //--- barriers are now the EA's own SL/TP, so ~89% of bars RESOLVE and only the timeouts are Neutral.
   //--- Measured on SP500 H4: Buy 48.26% Sell 41.13% Neutral 10.61%. Neutral became the RAREST class,
   //--- and the correction dutifully started subsidising it - by tau*(log pB - log pN) = 1.20 logits at
   //--- the capped tau of 0.79. With no directional edge to overcome that (direction is closed at
   //--- best-of-999, p=1.0000), the model took the free lunch: OOS recall Buy:1% Sell:0% Neutral:100%,
   //--- softmax saturated at spread 0.9993, and the first-layer weight block froze at 0.000% dW/W while
   //--- the head kept twitching. The anti-collapse mechanism WAS the collapse.
   //---
   //--- Neutral is not a class worth predicting here - it is the ABSTAIN outcome, and abstention is
   //--- already owned by a better mechanism: m_dirConfThreshold, refitted every era on the held-out
   //--- calibration band against a coverage floor and the measured break-even. Subsidising the abstain
   //--- class does the same job twice and spends the entire correction suppressing the only decisions
   //--- that can make money. What DOES deserve correcting is Buy vs Sell: a trending symbol resolves
   //--- more long barriers than short ones, and left uncorrected the model inherits that drift as a
   //--- standing directional bias. Here that is log(0.4826) - log(0.4113) = 0.16, so the offsets are
   //--- tiny - which is the correct answer, not a broken one. The two classes were already balanced;
   //--- all the old spread of 1.52 ever described was how rare a timeout is.
   //---
   //--- Centred on the midpoint of the two so the pair is corrected against EACH OTHER and Neutral sits
   //--- at zero. Softmax is shift-invariant, so only the differences matter: Neutral now sits within
   //--- tau*0.08 of both trading classes instead of 1.20 above them.
   double mid = 0.5 * (lb + ls);
   double spread = MathAbs(lb - ls);
   double tauEff = m_logitAdjustTau;
   if(spread > 0.0)
     {
      double cap = LOGIT_ADJUST_MAX_RANGE_FRACTION * CLASS_LOGIT_SCALE / spread;
      if(tauEff > cap)
         tauEff = cap;
     }
   if(!m_logitAdjustLogged)
     {
      m_logitAdjustLogged = true;
      //--- Reports BOTH spreads on purpose. The Buy-vs-Sell one is what is actually applied; the
      //--- all-three one is what the old code applied, and printing them side by side is what makes it
      //--- visible when a label set has drifted so far that the abstain class is the rare one.
      Print(ID + ": logit adjustment - measured priors Buy " + DoubleToString(m_priorBuy * 100.0, 2) +
            "% Sell " + DoubleToString(m_priorSell * 100.0, 2) + "% Neutral " +
            DoubleToString(m_priorNeutral * 100.0, 2) + "% | APPLIED across Buy/Sell only, log-prior"
            " spread " + DoubleToString(spread, 2) + " (all three would be " +
            DoubleToString(MathMax(lb, MathMax(ls, lnn)) - MathMin(lb, MathMin(ls, lnn)), 2) +
            "; Neutral is the ABSTAIN outcome and is never subsidised - m_dirConfThreshold owns"
            " abstention)" +
            (m_priorNeutral < m_priorBuy && m_priorNeutral < m_priorSell
             ? " | note: Neutral is the RAREST class here, so the pre-2026-08-16 all-three form would"
               " have BOOSTED it by " +
               DoubleToString(tauEff * (MathMax(lb, ls) - lnn), 2) + " logits"
             : "") +
            " against a logit range of " +
            DoubleToString(CLASS_LOGIT_SCALE, 1) + " | tau " + DoubleToString(m_logitAdjustTau, 2) +
            (tauEff < m_logitAdjustTau
             ? " CAPPED to " + DoubleToString(tauEff, 2) + " (uncapped it would consume " +
               DoubleToString(100.0 * spread * m_logitAdjustTau / CLASS_LOGIT_SCALE, 0) +
               "% of the range and saturate the head)"
             : " (uncapped - within budget)"));
     }
   //--- ORDERED to match the output layer: [0]=Buy, [1]=Sell, [2]=Neutral - the order
   //--- BuildFreshTopology emits and the order the softmax gradient reads (AI\Network.mqh).
   double offsets[3];
   offsets[0] = tauEff * (lb - mid);
   offsets[1] = tauEff * (ls - mid);
   offsets[2] = 0.0;   // ABSTAIN class - never subsidised; see the block above
   Net.SetLogitAdjustment(offsets);
  }
//+------------------------------------------------------------------+
//| Converts a double to ENUM_SIGNAL.                                 |
//| 3-output (softmax classification) case: dPrevSignal's *sign*      |
//| alone already encodes the argmax-selected class (+prob for Buy,   |
//| -prob for Sell, exactly 0.0 for Neutral - see Train()/            |
//| RefreshLatestSignal()), so classification here is pure argmax:    |
//| whichever class the network actually picked, full stop. No        |
//| magnitude threshold is applied - confidence magnitude is a         |
//| separate concern, already exposed via AIConfidence()/              |
//| SignedAIConfidence() (MathAbs(dPrevSignal)/dPrevSignal) for the    |
//| signal engine's own confidence-weighted filters/lot sizing/SLTP,   |
//| so this keeps "which class" and "how confident" decoupled.         |
//| 1-output (tanh regression) case: unrelated network shape, keeps    |
//| the original 0.50 magnitude cutoff as a genuine confidence gate.   |
//+------------------------------------------------------------------+
ENUM_SIGNAL CExpertSignalAIBase::DoubleToSignal(double value)
  {
   value = NormalizeDouble(value, 2); // Round 'value' to two decimal places
   if(value < -1.0 || value > 1.0)
      return Undefine; // out of range, e.g. the -2 "not yet studied" sentinel
   if(m_outputNeuronsCount == 3)
     {
      if(value > 0.0)
         return Buy;
      if(value < 0.0)
         return Sell;
      return Neutral;
     }
   if(value > 0.50)
      return Buy;
   if(value < -0.50)
      return Sell;
   return Neutral;
  }
//+------------------------------------------------------------------+
//| Throttled, SIDE-EFFECT-FREE forward of the current decision bar,  |
//| for display only (the HUD member lines and the prospective vote). |
//|                                                                   |
//| The reference library kept its training label honest by simply    |
//| printing the last training sample's outputs - but a shuffled pass-2|
//| sample is a random historical bar, and what the user tracks is the|
//| model's opinion of NOW under the weights of NOW. So this asks the |
//| exact question the live path asks (the window ending on bar 1, the|
//| newest CLOSED bar - see RefreshLatestSignal for why not bar 0) and |
//| touches NOTHING the trading or training paths read:               |
//|  - dPrevSignal, the NMS state, the refresh tallies, dtStudied and |
//|    m_lastBarTime all stay untouched - RefreshLatestSignal is NOT  |
//|    reusable here precisely because it writes all of them;         |
//|  - batch-norm running statistics are bracketed frozen/restored    |
//|    (GetBatchNormFrozen), because an unfrozen forward ADVANCES them|
//|    - hundreds of display reads per era would otherwise retrain the|
//|    normalization on one bar's window; restore-not-unfreeze because|
//|    pass 3 holds them frozen across its whole scan and a display   |
//|    tick landing between its chunks must not unfreeze mid-scan;    |
//|  - the LSTM is safe by construction: h_{-1}/c_{-1} are zeroed per |
//|    forward (see AI\Impl\NeuronOCLLSTM.mqh), nothing leaks between |
//|    samples;                                                       |
//|  - TempData is the shared scratch every consumer rebuilds before  |
//|    use, and this builds/forwards/reads it atomically.             |
//|                                                                   |
//| It forwards Net - the LEARNER - not the shadow: the shadow is what|
//| trades, but EnsureShadowNet()'s first call clones a full net      |
//| through a temp file, a side effect a display routine must never   |
//| trigger, and during training (the whole use case) the shadow lags |
//| the learner by construction. Post-convergence Net holds the       |
//| converged weights and online learning keeps updating it, so the   |
//| line stays honest there too.                                      |
//|                                                                   |
//| Throttle: a real forward at most every DISPLAY_FWD_MIN_MS, or     |
//| DISPLAY_FWD_ERA_MS after an era boundary (weights AND tier money  |
//| just moved, the cached read is priced in a dead regime). Between  |
//| refreshes the cached m_dispProbs/m_dispSignal serve every caller, |
//| so the 500ms timer costs nothing extra. Failures keep the last    |
//| good read on display (stale-by-seconds beats blank) but stamp the |
//| attempt, so a broken window retries at throttle pace, not 2/sec.  |
//+------------------------------------------------------------------+
//--- (uint) so the throttle comparison is unsigned-vs-unsigned: age is a uint tick delta, and
//--- the ternary picking between these is a runtime expression the compiler cannot constant-
//--- fold, so bare int literals here drew a sign-mismatch warning (reported 2026-08-19).
#define DISPLAY_FWD_MIN_MS ((uint)4000)
#define DISPLAY_FWD_ERA_MS ((uint)1000)
bool CExpertSignalAIBase::DisplayInference(void)
  {
   //--- Meta head consumes fired candidates, not a bare bar window - a candidate-less forward is
   //--- width-mismatched against its input layer. Same guard as RefreshConvergedSignal.
   if(IsMetaTarget())
      return false;
   if(CheckPointer(Net) == POINTER_INVALID)
      return false;
   if(m_outputNeuronsCount != 1 && m_outputNeuronsCount != 3)
      return false;
   uint now = GetTickCount();
   uint age = now - m_dispStamp;   // unsigned subtraction survives the 49-day wrap
   bool eraMoved = ((long)m_eraCount != m_dispEra);
   if(m_dispStamp != 0 && age < (eraMoved ? DISPLAY_FWD_ERA_MS : DISPLAY_FWD_MIN_MS))
      return m_dispValid;          // serve the cache (or keep failing quietly) until the throttle opens
   m_dispStamp = now;
   if(!BuildFeatureWindow(1))
      return m_dispValid;          // window not buildable (warm-up, indicator hole): keep the last read
   //--- Save/restore, NOT set/clear - see the header. Frozen, this forward is a pure function.
   bool bnWasFrozen = Net.GetBatchNormFrozen();
   if(!bnWasFrozen)
      Net.SetBatchNormFrozen(true);
   bool fwdOk = Net.feedForward(TempData);
   if(fwdOk)
      Net.getResults(TempData);
   if(!bnWasFrozen)
      Net.SetBatchNormFrozen(false);
   if(!fwdOk)
      return m_dispValid;
   if(m_outputNeuronsCount == 1)
     {
      double v = TempData.At(0);
      if(!MathIsValidNumber(v))
         return m_dispValid;       // NaN net: keep the last finite read, the NaN latch reports elsewhere
      m_dispProbs[0] = v;
      m_dispProbs[1] = 0.0;
      m_dispProbs[2] = 0.0;
      m_dispSignal   = v;
     }
   else
     {
      //--- Same two calls, same order, as the live decision in RefreshLatestSignal: softmax INTO
      //--- TempData, then the strict-majority read. On non-finite logits the softmax returns 0
      //--- WITHOUT normalizing TempData - the finiteness check below is what keeps raw NaN logits
      //--- from being displayed as probabilities.
      ApplyClassificationSoftmax();
      double p0 = TempData.At(0), p1 = TempData.At(1), p2 = TempData.At(2);
      if(!MathIsValidNumber(p0) || !MathIsValidNumber(p1) || !MathIsValidNumber(p2))
         return m_dispValid;
      m_dispProbs[0] = p0;   // Buy
      m_dispProbs[1] = p1;   // Sell
      m_dispProbs[2] = p2;   // Neutral
      m_dispSignal   = AdjustedSignalFromSoftmax();
     }
   m_dispEra   = (long)m_eraCount;
   m_dispValid = true;
   return true;
  }
#endif // WARRIOR_AIBASE_INFERENCE_MQH