Guard template selector confidence by support kind

UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistTraining/HyperTwistCoachDashboardWidget.cpp

@@ -756,8 +756,8 @@ namespace HyperTwistCoachDashboardWidgetInternal
 	}
 
 	constexpr int32 TemplateLaunchAnalyticsTieBreakMinimumOutcomeRuns = 2;
-	constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumTieBreakOutcomeRuns = 3;
-	constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumPrimaryOutcomeRuns = 2;
+	constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumSplitConfidenceOutcomeRuns = 3;
+	constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumTemplateSignalOutcomeRuns = 2;
 	constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumScoreDelta = 2;
 	constexpr float TemplateLaunchSelectionTieBreakConfidenceMinimumSuccessDelta = 0.05f;
 	constexpr float TemplateLaunchSelectionTieBreakConfidenceMinimumMistakeImprovement = 0.5f;
@@ -847,29 +847,37 @@ namespace HyperTwistCoachDashboardWidgetInternal
 			return 0;
 		}
 
-		const FHyperTwistTrainingTemplateLaunchSelectionScopedStats& TieBreakStats =
-			TemplateStats->DashboardAnalyticsTieBreakStats;
-		const FHyperTwistTrainingTemplateLaunchSelectionScopedStats& PrimaryStats =
-			TemplateStats->DashboardPrimarySelectionStats;
-		const int32 TieBreakOutcomeRuns = CountTemplateLaunchSelectionOutcomeRuns(&TieBreakStats);
-		const int32 PrimaryOutcomeRuns = CountTemplateLaunchSelectionOutcomeRuns(&PrimaryStats);
-		if (TieBreakOutcomeRuns < TemplateLaunchSelectionTieBreakConfidenceMinimumTieBreakOutcomeRuns
-			|| PrimaryOutcomeRuns < TemplateLaunchSelectionTieBreakConfidenceMinimumPrimaryOutcomeRuns)
+		const FHyperTwistTrainingTemplateLaunchSelectionScopedStats& SplitConfidenceStats =
+			TemplateStats->DashboardAnalyticsTieBreakSelectorSplitConfidenceStats;
+		const FHyperTwistTrainingTemplateLaunchSelectionScopedStats& TemplateSignalStats =
+			TemplateStats->DashboardAnalyticsTieBreakTemplateScopedSignalStats;
+		const int32 SplitConfidenceOutcomeRuns =
+			CountTemplateLaunchSelectionOutcomeRuns(&SplitConfidenceStats);
+		const int32 TemplateSignalOutcomeRuns =
+			CountTemplateLaunchSelectionOutcomeRuns(&TemplateSignalStats);
+		if (SplitConfidenceOutcomeRuns
+				< TemplateLaunchSelectionTieBreakConfidenceMinimumSplitConfidenceOutcomeRuns
+			|| TemplateSignalOutcomeRuns
+				< TemplateLaunchSelectionTieBreakConfidenceMinimumTemplateSignalOutcomeRuns)
 		{
 			return 0;
 		}
 
-		const int32 TieBreakScore = BuildTemplateLaunchSelectionLaneAnalyticsScore(&TieBreakStats);
-		const int32 PrimaryScore = BuildTemplateLaunchSelectionLaneAnalyticsScore(&PrimaryStats);
-		const int32 ScoreDelta = TieBreakScore - PrimaryScore;
+		const int32 SplitConfidenceScore =
+			BuildTemplateLaunchSelectionLaneAnalyticsScore(&SplitConfidenceStats);
+		const int32 TemplateSignalScore =
+			BuildTemplateLaunchSelectionLaneAnalyticsScore(&TemplateSignalStats);
+		const int32 ScoreDelta = SplitConfidenceScore - TemplateSignalScore;
 		if (ScoreDelta <= 0)
 		{
 			return 0;
 		}
 
-		const float SuccessDelta = TieBreakStats.OverallSuccessRate - PrimaryStats.OverallSuccessRate;
+		const float SuccessDelta =
+			SplitConfidenceStats.OverallSuccessRate - TemplateSignalStats.OverallSuccessRate;
 		const float MistakeImprovement =
-			PrimaryStats.AverageMistakesPerOutcomeRun - TieBreakStats.AverageMistakesPerOutcomeRun;
+			TemplateSignalStats.AverageMistakesPerOutcomeRun
+			- SplitConfidenceStats.AverageMistakesPerOutcomeRun;
 		if (ScoreDelta < TemplateLaunchSelectionTieBreakConfidenceMinimumScoreDelta
 			&& SuccessDelta < TemplateLaunchSelectionTieBreakConfidenceMinimumSuccessDelta
 			&& MistakeImprovement < TemplateLaunchSelectionTieBreakConfidenceMinimumMistakeImprovement)

docs/arch/HYPERTWIST_PHASE4_TEMPLATE_LAUNCH_SUPPORT_KIND_CONFIDENCE_GUARD_PACKET_2026-05-07.md

@@ -0,0 +1,92 @@
+# HyperTwist Phase 4 template launch support-kind confidence guard packet
+
+Created on `2026-05-07`
+
+Status:
+
+- first-party HyperTwist packet
+- bounded Phase `4` dashboard selector refinement slice
+
+## Purpose
+
+This packet replaces the selector-split confidence bonus guard so it is backed by repo-derived support-kind buckets instead of the broader analytics-tie-break lane as a whole.
+
+The open tasks were:
+
+- stop inferring selector-split confidence from mixed analytics tie-break history
+- require the stored `SelectorSplitConfidence` bucket itself to have enough outcome history before it can strengthen a future tied selection
+- compare that bucket directly against the template's `TemplateScopedOutcomeSignal` tie-break bucket rather than against non-tie-break dashboard priority history
+
+It is not:
+
+- a new repository aggregation packet
+- another analytics provenance packet
+- a change to primary dashboard template ranking rules
+- a new dashboard surface packet
+
+## Scope
+
+Bounded lane:
+
+- retarget the dashboard selector confidence guard to repo-backed support-kind buckets already present on template launch scoped stats
+- require enough stored outcome runs on the `SelectorSplitConfidence` bucket and the `TemplateScopedOutcomeSignal` bucket before awarding the extra confidence point
+- keep the same small confidence bonus and the same bounded score/success/mistake thresholds
+- leave the rest of the overall analytics tie-break path unchanged
+
+Out of scope:
+
+- changing support-kind aggregation itself
+- changing the existing template-scoped analytics tie-break score
+- changing selection provenance payloads
+- adding a new dashboard explanation surface
+
+## Why this was the right next packet
+
+The previous packet made support-kind split outcomes visible in repo-backed template analytics.
+
+That immediately exposed a selector gap:
+
+- the dashboard could now show support-kind split performance
+- but the selector confidence bonus still looked at the broader analytics tie-break lane instead of the support-kind split itself
+- which meant mixed tie-break history could still award selector-split confidence even if the explicit `SelectorSplitConfidence` bucket had not independently proven itself
+
+So the next honest move was:
+
+- keep the bonus small
+- keep it attached only to tied-template analytics selection
+- but make it answer to the actual support-kind split that now exists in stored analytics
+
+## What landed
+
+Primary code changes:
+
+- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistTraining/HyperTwistCoachDashboardWidget.cpp`
+  - retargeted `BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(...)` to compare `DashboardAnalyticsTieBreakSelectorSplitConfidenceStats` against `DashboardAnalyticsTieBreakTemplateScopedSignalStats`
+  - requires enough stored outcome runs on both support-kind buckets before awarding confidence
+  - preserves the existing bounded score delta, success delta, and mistake-improvement thresholds
+  - leaves the broader equal-priority analytics tie-break path unchanged outside this confidence source swap
+
+## Product effect
+
+Dashboard selector-split confidence is now grounded in the right evidence:
+
+- a template only earns the extra selector-split confidence point when its explicit `SelectorSplitConfidence` history has enough outcomes
+- that bucket must outperform the template's own `TemplateScopedOutcomeSignal` tie-break history, not merely the broader mixed analytics lane
+- sparse or mixed tie-break history no longer overstates selector-split confidence
+- tied-template selection remains conservative and repo-backed
+
+## Acceptance criteria
+
+- selector-split confidence bonus uses repo-backed support-kind buckets
+- bonus only applies when the `SelectorSplitConfidence` bucket has enough outcome history
+- bonus only applies when the `SelectorSplitConfidence` bucket outperforms the `TemplateScopedOutcomeSignal` bucket on meaningful stored signal
+- full product build succeeds
+
+## Validation checklist
+
+1. build `UnrealHyperTwist.sln` / `UnrealHyperTwistEditor`
+2. confirm selector confidence bonus now reads support-kind split stats instead of the broader analytics tie-break lane
+3. confirm sparse or underperforming `SelectorSplitConfidence` history does not award the extra confidence point
+4. confirm the rest of the tied-template analytics selection path remains unchanged
+
+That is the packet.