Guard template confidence by comparable peer evidence
This commit is contained in:
axiomlogicnexus
parent
f9fbc23216
commit
badbd4a082
2 changed files with 180 additions and 1 deletions
|
|
@ -121,6 +121,14 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats
|
||||
);
|
||||
|
||||
bool HasTemplateLaunchComparablePeerEvidenceForConfidenceBonus(
|
||||
const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats,
|
||||
int32* OutComparablePeerOutcomeRuns = nullptr,
|
||||
int32* OutSplitConfidenceOutcomeRuns = nullptr,
|
||||
float* OutComparablePeerOutcomeShare = nullptr,
|
||||
int32* OutPeerNoEvidenceOutcomeRuns = nullptr
|
||||
);
|
||||
|
||||
int32 BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(
|
||||
const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats
|
||||
);
|
||||
|
|
@ -930,6 +938,8 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumScoreDelta = 2;
|
||||
constexpr float TemplateLaunchSelectionTieBreakConfidenceMinimumSuccessDelta = 0.05f;
|
||||
constexpr float TemplateLaunchSelectionTieBreakConfidenceMinimumMistakeImprovement = 0.5f;
|
||||
constexpr int32 TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeRuns = 2;
|
||||
constexpr float TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeShare = 0.40f;
|
||||
|
||||
int32 CountTemplateLaunchOutcomeRuns(const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats)
|
||||
{
|
||||
|
|
@ -1007,6 +1017,73 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
);
|
||||
}
|
||||
|
||||
bool HasTemplateLaunchComparablePeerEvidenceForConfidenceBonus(
|
||||
const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats,
|
||||
int32* OutComparablePeerOutcomeRuns,
|
||||
int32* OutSplitConfidenceOutcomeRuns,
|
||||
float* OutComparablePeerOutcomeShare,
|
||||
int32* OutPeerNoEvidenceOutcomeRuns
|
||||
)
|
||||
{
|
||||
const FHyperTwistTrainingTemplateLaunchSelectionScopedStats* SplitConfidenceStats = nullptr;
|
||||
const FHyperTwistTrainingTemplateLaunchSelectionScopedStats* PeerNoEvidenceStats = nullptr;
|
||||
const FHyperTwistTrainingTemplateLaunchSelectionScopedStats* PeerWeakerStats = nullptr;
|
||||
const FHyperTwistTrainingTemplateLaunchSelectionScopedStats* PeerMarginalStats = nullptr;
|
||||
const FHyperTwistTrainingTemplateLaunchSelectionScopedStats* PeerClearedStats = nullptr;
|
||||
if (TemplateStats != nullptr && TemplateStats->IsStructurallyValid())
|
||||
{
|
||||
SplitConfidenceStats =
|
||||
&TemplateStats->DashboardAnalyticsTieBreakSelectorSplitConfidenceStats;
|
||||
PeerNoEvidenceStats =
|
||||
&TemplateStats
|
||||
->DashboardAnalyticsTieBreakSelectorSplitConfidencePeerLackedComparableEvidenceStats;
|
||||
PeerWeakerStats =
|
||||
&TemplateStats
|
||||
->DashboardAnalyticsTieBreakSelectorSplitConfidencePeerFailedWeakerSupportKindStats;
|
||||
PeerMarginalStats =
|
||||
&TemplateStats
|
||||
->DashboardAnalyticsTieBreakSelectorSplitConfidencePeerFailedMarginalSupportKindStats;
|
||||
PeerClearedStats =
|
||||
&TemplateStats
|
||||
->DashboardAnalyticsTieBreakSelectorSplitConfidencePeerClearedComparableGuardStats;
|
||||
}
|
||||
|
||||
const int32 ComparablePeerOutcomeRuns =
|
||||
CountTemplateLaunchSelectionOutcomeRuns(PeerWeakerStats)
|
||||
+ CountTemplateLaunchSelectionOutcomeRuns(PeerMarginalStats)
|
||||
+ CountTemplateLaunchSelectionOutcomeRuns(PeerClearedStats);
|
||||
const int32 SplitConfidenceOutcomeRuns =
|
||||
CountTemplateLaunchSelectionOutcomeRuns(SplitConfidenceStats);
|
||||
const int32 PeerNoEvidenceOutcomeRuns =
|
||||
CountTemplateLaunchSelectionOutcomeRuns(PeerNoEvidenceStats);
|
||||
const float ComparablePeerOutcomeShare = SplitConfidenceOutcomeRuns > 0
|
||||
? static_cast<float>(ComparablePeerOutcomeRuns)
|
||||
/ static_cast<float>(SplitConfidenceOutcomeRuns)
|
||||
: 0.0f;
|
||||
|
||||
if (OutComparablePeerOutcomeRuns != nullptr)
|
||||
{
|
||||
*OutComparablePeerOutcomeRuns = ComparablePeerOutcomeRuns;
|
||||
}
|
||||
if (OutSplitConfidenceOutcomeRuns != nullptr)
|
||||
{
|
||||
*OutSplitConfidenceOutcomeRuns = SplitConfidenceOutcomeRuns;
|
||||
}
|
||||
if (OutComparablePeerOutcomeShare != nullptr)
|
||||
{
|
||||
*OutComparablePeerOutcomeShare = ComparablePeerOutcomeShare;
|
||||
}
|
||||
if (OutPeerNoEvidenceOutcomeRuns != nullptr)
|
||||
{
|
||||
*OutPeerNoEvidenceOutcomeRuns = PeerNoEvidenceOutcomeRuns;
|
||||
}
|
||||
|
||||
return ComparablePeerOutcomeRuns
|
||||
>= TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeRuns
|
||||
&& ComparablePeerOutcomeShare
|
||||
>= TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeShare;
|
||||
}
|
||||
|
||||
int32 BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(
|
||||
const FHyperTwistTrainingTemplateLaunchScopedStats* TemplateStats
|
||||
)
|
||||
|
|
@ -1053,6 +1130,10 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
{
|
||||
return 0;
|
||||
}
|
||||
if (!HasTemplateLaunchComparablePeerEvidenceForConfidenceBonus(TemplateStats))
|
||||
{
|
||||
return 0;
|
||||
}
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
|
@ -1094,6 +1175,11 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
return EHyperTwistTrainingTemplateLaunchSelectionPeerSupportKindDisposition::
|
||||
PeerFailedGuardOnWeakerSupportKindOutcomes;
|
||||
}
|
||||
if (!HasTemplateLaunchComparablePeerEvidenceForConfidenceBonus(TemplateStats))
|
||||
{
|
||||
return EHyperTwistTrainingTemplateLaunchSelectionPeerSupportKindDisposition::
|
||||
PeerLackedComparableEvidence;
|
||||
}
|
||||
|
||||
if (BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(TemplateStats) <= 0)
|
||||
{
|
||||
|
|
@ -1166,6 +1252,32 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
const float MistakeImprovement =
|
||||
TemplateSignalStats.AverageMistakesPerOutcomeRun
|
||||
- SplitConfidenceStats.AverageMistakesPerOutcomeRun;
|
||||
int32 ComparablePeerOutcomeRuns = 0;
|
||||
int32 SplitConfidenceSelectionOutcomeRuns = 0;
|
||||
int32 PeerNoEvidenceOutcomeRuns = 0;
|
||||
float ComparablePeerOutcomeShare = 0.0f;
|
||||
const bool bHasComparablePeerEvidence =
|
||||
HasTemplateLaunchComparablePeerEvidenceForConfidenceBonus(
|
||||
TemplateStats,
|
||||
&ComparablePeerOutcomeRuns,
|
||||
&SplitConfidenceSelectionOutcomeRuns,
|
||||
&ComparablePeerOutcomeShare,
|
||||
&PeerNoEvidenceOutcomeRuns
|
||||
);
|
||||
if (!bHasComparablePeerEvidence)
|
||||
{
|
||||
return FString::Printf(
|
||||
TEXT("%s lacked enough selector-split wins against peers with comparable support-kind evidence to clear the selector-split confidence guard: comparable-peer outcomes %d of %d split-confidence outcomes (share %.2f, required at least %d at %.2f) | peer-no-evidence %d | %s"),
|
||||
ResolvedSubjectLabel,
|
||||
ComparablePeerOutcomeRuns,
|
||||
SplitConfidenceSelectionOutcomeRuns,
|
||||
ComparablePeerOutcomeShare,
|
||||
TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeRuns,
|
||||
TemplateLaunchSelectionTieBreakConfidenceMinimumComparablePeerOutcomeShare,
|
||||
PeerNoEvidenceOutcomeRuns,
|
||||
*SignalFragment
|
||||
);
|
||||
}
|
||||
if (BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(TemplateStats) <= 0)
|
||||
{
|
||||
return FString::Printf(
|
||||
|
|
@ -1178,8 +1290,11 @@ namespace HyperTwistCoachDashboardWidgetInternal
|
|||
}
|
||||
|
||||
return FString::Printf(
|
||||
TEXT("%s split-confidence beat its template-signal bucket on stored support-kind outcomes and cleared the selector-split confidence guard: %s"),
|
||||
TEXT("%s split-confidence beat its template-signal bucket on stored support-kind outcomes and cleared the selector-split confidence guard with comparable-peer outcomes %d of %d split-confidence outcomes (share %.2f): %s"),
|
||||
ResolvedSubjectLabel,
|
||||
ComparablePeerOutcomeRuns,
|
||||
SplitConfidenceSelectionOutcomeRuns,
|
||||
ComparablePeerOutcomeShare,
|
||||
*SignalFragment
|
||||
);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -0,0 +1,64 @@
|
|||
# HyperTwist Phase 4 template launch comparable-peer confidence guard packet
|
||||
|
||||
Created on `2026-05-07`
|
||||
|
||||
Status:
|
||||
|
||||
- first-party HyperTwist packet
|
||||
- bounded Phase `4` dashboard selector confidence slice
|
||||
|
||||
## Purpose
|
||||
|
||||
This packet starts consuming the new peer-support disposition analytics in actual selector policy. The selector-split confidence bonus now requires enough wins against peers that had at least some comparable support-kind evidence, so stored confidence does not keep reinforcing itself mainly by beating peer lanes that had no comparable evidence at all.
|
||||
|
||||
## Scope
|
||||
|
||||
Bounded lane:
|
||||
|
||||
- keep the existing support-kind score, success, and mistake delta guard intact
|
||||
- add a comparable-peer-evidence guard on top of it using the repo-backed peer disposition buckets
|
||||
- require both:
|
||||
- a minimum count of selector-split-confidence outcomes against peers with comparable evidence
|
||||
- a minimum share of the template's selector-split-confidence outcome history coming from those comparable-evidence peers
|
||||
- align the existing support-kind audit clause with the stricter guard so rationale stays honest
|
||||
|
||||
Out of scope:
|
||||
|
||||
- changing repo aggregation shape
|
||||
- widening persisted provenance schema again
|
||||
- adding a new dashboard panel
|
||||
- feeding peer disposition directly into primary template scoring outside the selector-split confidence bonus
|
||||
|
||||
## What landed
|
||||
|
||||
Primary code changes:
|
||||
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistTraining/HyperTwistCoachDashboardWidget.cpp`
|
||||
- added a helper that measures comparable-peer selector-split-confidence evidence from the repo-backed peer disposition buckets
|
||||
- `BuildTemplateLaunchAnalyticsTieBreakConfidenceBonus(...)` now refuses the bonus unless comparable-peer outcome history is deep enough and large enough as a share of the template's selector-split-confidence history
|
||||
- peer support-kind disposition resolution now treats insufficient comparable-peer evidence as `PeerLackedComparableEvidence`
|
||||
- support-kind audit text now explains the comparable-peer evidence requirement and shows counts/share when that guard fails or clears
|
||||
|
||||
## Product effect
|
||||
|
||||
Template selector confidence is more conservative:
|
||||
|
||||
- selector-split confidence reinforcement still requires the template's split-confidence bucket to beat its template-signal bucket
|
||||
- but that is no longer enough by itself
|
||||
- the template must also show enough split-confidence wins against peers that brought some comparable evidence, rather than relying mostly on peer-no-evidence lanes
|
||||
|
||||
## Acceptance criteria
|
||||
|
||||
- selector-split confidence bonus requires minimum comparable-peer outcome count and share
|
||||
- insufficient comparable-peer evidence suppresses the bonus even when the support-kind split beats the template-signal bucket
|
||||
- peer disposition and support-kind audit text stay aligned with the stricter guard
|
||||
- full product build succeeds
|
||||
|
||||
## Validation checklist
|
||||
|
||||
1. build `UnrealHyperTwist.sln` / `UnrealHyperTwistEditor`
|
||||
2. confirm selector-split confidence bonus now consults the repo-backed peer disposition buckets
|
||||
3. confirm insufficient comparable-peer evidence suppresses the bonus
|
||||
4. confirm support-kind audit text names the comparable-peer count/share requirement when relevant
|
||||
|
||||
That is the packet.
|
||||
Loading…
Add table
Reference in a new issue