feat(capability): add task-aware model evaluation evidence
This commit is contained in:
@@ -148,6 +148,9 @@ func SelectCognitiveBinding(requirement capability.CognitiveRequirement, candida
|
||||
if candidate.Evaluation != nil && candidate.Evaluation.Result == capability.EvaluationUnsupported {
|
||||
continue
|
||||
}
|
||||
if requirement.TaskProfileID != "" && (candidate.TaskEvaluation == nil || candidate.TaskEvaluation.ProfileID != requirement.TaskProfileID || candidate.TaskEvaluation.Result != capability.EvaluationSupported) {
|
||||
continue
|
||||
}
|
||||
rating := RateCognitiveBinding(candidate)
|
||||
candidate.Rating = &rating
|
||||
if requirement.MaxCostMicros != nil {
|
||||
@@ -171,6 +174,19 @@ func SelectCognitiveBinding(requirement capability.CognitiveRequirement, candida
|
||||
// evidence. Provider kind, endpoint name and hardware are intentionally not
|
||||
// inputs to the score.
|
||||
func RateCognitiveBinding(candidate capability.ResolvedCognitiveBinding) capability.ModelRating {
|
||||
if candidate.TaskEvaluation != nil {
|
||||
evaluation := candidate.TaskEvaluation
|
||||
if evaluation.Result != capability.EvaluationSupported {
|
||||
return capability.ModelRating{TaskProfileID: evaluation.ProfileID, Category: capability.ModelRatingUnrated}
|
||||
}
|
||||
category := capability.ModelRatingBalanced
|
||||
if evaluation.OverallScoreMillis >= 850 && evaluation.QualityScoreMillis >= 800 {
|
||||
category = capability.ModelRatingQuality
|
||||
} else if evaluation.LatencyScoreMillis >= 800 && evaluation.OverallScoreMillis >= 700 {
|
||||
category = capability.ModelRatingFast
|
||||
}
|
||||
return capability.ModelRating{TaskProfileID: evaluation.ProfileID, TaskScoreMillis: evaluation.OverallScoreMillis, Category: category, OverallScoreMillis: evaluation.OverallScoreMillis, QualityScoreMillis: evaluation.QualityScoreMillis, ReliabilityScoreMillis: evaluation.ReliabilityScoreMillis, LatencyScoreMillis: evaluation.LatencyScoreMillis, EvidenceScoreMillis: 1000}
|
||||
}
|
||||
if candidate.Evaluation == nil || candidate.Evaluation.Result != capability.EvaluationSupported || candidate.Evaluation.QualityScoreMillis == nil || candidate.Evaluation.CriticalFailureMillis == nil {
|
||||
return capability.ModelRating{Category: capability.ModelRatingUnrated}
|
||||
}
|
||||
|
||||
@@ -0,0 +1,106 @@
|
||||
package domain
|
||||
|
||||
import (
|
||||
"errors"
|
||||
|
||||
"git.cin.su/m/Assistant/contracts/artifact"
|
||||
"git.cin.su/m/Assistant/contracts/capability"
|
||||
)
|
||||
|
||||
var ErrInvalidTaskEvaluation = errors.New("invalid task evaluation")
|
||||
|
||||
func BuildAdditionalTaskEvaluationProfile(parent capability.TaskEvaluationProfile, request capability.TaskEvaluationExpansionRequest, profileID capability.TaskEvaluationProfileID, revision int64) (capability.TaskEvaluationProfile, error) {
|
||||
if err := parent.Validate(); err != nil || parent.Stage != capability.TaskEvaluationGeneral || request.ParentProfileID != parent.ID || request.CapabilityID != parent.CapabilityID || request.TriggerEvidenceRef == "" || len(request.NewCaseIDs) == 0 || profileID == "" || revision < 1 || request.BenchmarkVersion == "" {
|
||||
return capability.TaskEvaluationProfile{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
known := make(map[string]struct{}, len(parent.RequiredCaseIDs))
|
||||
for _, caseID := range parent.RequiredCaseIDs {
|
||||
known[caseID] = struct{}{}
|
||||
}
|
||||
seen := make(map[string]struct{}, len(request.NewCaseIDs))
|
||||
for _, caseID := range request.NewCaseIDs {
|
||||
if caseID == "" {
|
||||
return capability.TaskEvaluationProfile{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
if _, exists := known[caseID]; exists {
|
||||
return capability.TaskEvaluationProfile{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
if _, exists := seen[caseID]; exists {
|
||||
return capability.TaskEvaluationProfile{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
seen[caseID] = struct{}{}
|
||||
}
|
||||
return capability.TaskEvaluationProfile{ID: profileID, Revision: revision, CapabilityID: parent.CapabilityID, Stage: capability.TaskEvaluationAdditional, ParentProfileID: parent.ID, TriggerEvidenceRef: request.TriggerEvidenceRef, BenchmarkVersion: request.BenchmarkVersion, RequiredCaseIDs: append([]string(nil), request.NewCaseIDs...), QualityWeightMillis: parent.QualityWeightMillis, ReliabilityWeightMillis: parent.ReliabilityWeightMillis, LatencyWeightMillis: parent.LatencyWeightMillis}, nil
|
||||
}
|
||||
|
||||
// BuildTaskEvaluation aggregates independently observed case results. It is
|
||||
// intentionally deterministic: no provider name, hardware label or caller
|
||||
// supplied aggregate score can influence the result.
|
||||
func BuildTaskEvaluation(profile capability.TaskEvaluationProfile, evaluationID capability.TaskEvaluationID, bindingID capability.CapabilityBindingID, bindingRevision int64, capabilityID capability.CapabilityID, observations []capability.TaskCaseObservation, artifactID artifact.ArtifactID, observedAtUnixMillis int64) (capability.TaskEvaluation, error) {
|
||||
if err := profile.Validate(); err != nil || evaluationID == "" || bindingID == "" || bindingRevision < 1 || capabilityID == "" || capabilityID != profile.CapabilityID || artifactID == "" || observedAtUnixMillis < 1 || len(observations) != len(profile.RequiredCaseIDs) {
|
||||
return capability.TaskEvaluation{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
required := make(map[string]struct{}, len(profile.RequiredCaseIDs))
|
||||
for _, caseID := range profile.RequiredCaseIDs {
|
||||
required[caseID] = struct{}{}
|
||||
}
|
||||
seen := make(map[string]struct{}, len(observations))
|
||||
qualityTotal, latencyTotal, passed, critical := int64(0), int64(0), int64(0), int64(0)
|
||||
unknown := false
|
||||
unsupported := false
|
||||
for _, observation := range observations {
|
||||
if _, ok := required[observation.CaseID]; !ok {
|
||||
return capability.TaskEvaluation{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
if _, ok := seen[observation.CaseID]; ok || observation.EvidenceArtifactRef == "" || observation.QualityScoreMillis < 0 || observation.QualityScoreMillis > 1000 || observation.LatencyMicros < 0 {
|
||||
return capability.TaskEvaluation{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
seen[observation.CaseID] = struct{}{}
|
||||
qualityTotal += observation.QualityScoreMillis
|
||||
latencyTotal += observation.LatencyMicros
|
||||
if observation.CriticalFailure {
|
||||
critical++
|
||||
}
|
||||
switch observation.Result {
|
||||
case capability.EvaluationSupported:
|
||||
passed++
|
||||
case capability.EvaluationUnsupported:
|
||||
unsupported = true
|
||||
case capability.EvaluationUnknown:
|
||||
unknown = true
|
||||
default:
|
||||
return capability.TaskEvaluation{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
}
|
||||
if len(seen) != len(required) {
|
||||
return capability.TaskEvaluation{}, ErrInvalidTaskEvaluation
|
||||
}
|
||||
count := int64(len(observations))
|
||||
quality := qualityTotal / count
|
||||
reliability := passed * 1000 / count
|
||||
latency := latencyScore(latencyTotal / count)
|
||||
criticalMillis := critical * 1000 / count
|
||||
overall := (quality*profile.QualityWeightMillis + reliability*profile.ReliabilityWeightMillis + latency*profile.LatencyWeightMillis) / 1000
|
||||
result := capability.EvaluationSupported
|
||||
if unknown {
|
||||
result = capability.EvaluationUnknown
|
||||
} else if unsupported {
|
||||
result = capability.EvaluationUnsupported
|
||||
}
|
||||
return capability.TaskEvaluation{ID: evaluationID, ProfileID: profile.ID, BindingID: bindingID, BindingRevision: bindingRevision, CapabilityID: capabilityID, Result: result, QualityScoreMillis: quality, ReliabilityScoreMillis: reliability, LatencyScoreMillis: latency, OverallScoreMillis: overall, CriticalFailureMillis: criticalMillis, BenchmarkVersion: profile.BenchmarkVersion, ArtifactID: artifactID, ObservedAtUnixMillis: observedAtUnixMillis}, nil
|
||||
}
|
||||
|
||||
func ValidateTaskEvaluation(value capability.TaskEvaluation) error {
|
||||
if value.ID == "" || value.ProfileID == "" || value.BindingID == "" || value.BindingRevision < 1 || value.CapabilityID == "" || value.BenchmarkVersion == "" || value.ArtifactID == "" || value.ObservedAtUnixMillis < 1 {
|
||||
return ErrInvalidTaskEvaluation
|
||||
}
|
||||
if value.Result != capability.EvaluationUnknown && value.Result != capability.EvaluationSupported && value.Result != capability.EvaluationUnsupported {
|
||||
return ErrInvalidTaskEvaluation
|
||||
}
|
||||
for _, score := range []int64{value.QualityScoreMillis, value.ReliabilityScoreMillis, value.LatencyScoreMillis, value.OverallScoreMillis, value.CriticalFailureMillis} {
|
||||
if score < 0 || score > 1000 {
|
||||
return ErrInvalidTaskEvaluation
|
||||
}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
@@ -0,0 +1,35 @@
|
||||
package domain
|
||||
|
||||
import (
|
||||
"testing"
|
||||
|
||||
"git.cin.su/m/Assistant/contracts/artifact"
|
||||
"git.cin.su/m/Assistant/contracts/capability"
|
||||
)
|
||||
|
||||
func TestBuildTaskEvaluationAggregatesOnlyProfileCases(t *testing.T) {
|
||||
profile := capability.TaskEvaluationProfile{ID: "research-v1", Revision: 1, CapabilityID: "creator-advisory", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "research-v1", RequiredCaseIDs: []string{"claim", "uncertainty"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
|
||||
value, err := BuildTaskEvaluation(profile, "evaluation-1", "binding-1", 2, "creator-advisory", []capability.TaskCaseObservation{
|
||||
{CaseID: "claim", Result: capability.EvaluationSupported, QualityScoreMillis: 900, LatencyMicros: 400_000, EvidenceArtifactRef: "artifact-case-1"},
|
||||
{CaseID: "uncertainty", Result: capability.EvaluationSupported, QualityScoreMillis: 800, LatencyMicros: 600_000, EvidenceArtifactRef: "artifact-case-2"},
|
||||
}, artifact.ArtifactID("artifact-summary"), 10)
|
||||
if err != nil || value.Result != capability.EvaluationSupported || value.OverallScoreMillis == 0 || value.ProfileID != profile.ID {
|
||||
t.Fatalf("evaluation=%+v err=%v", value, err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestBuildTaskEvaluationDoesNotPromoteUnknown(t *testing.T) {
|
||||
profile := capability.TaskEvaluationProfile{ID: "coding-v1", Revision: 1, CapabilityID: "creator-advisory", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "coding-v1", RequiredCaseIDs: []string{"compile"}, QualityWeightMillis: 700, ReliabilityWeightMillis: 200, LatencyWeightMillis: 100}
|
||||
value, err := BuildTaskEvaluation(profile, "evaluation-2", "binding-1", 2, "creator-advisory", []capability.TaskCaseObservation{{CaseID: "compile", Result: capability.EvaluationUnknown, QualityScoreMillis: 0, LatencyMicros: 0, EvidenceArtifactRef: "artifact-case"}}, artifact.ArtifactID("artifact-summary"), 11)
|
||||
if err != nil || value.Result != capability.EvaluationUnknown {
|
||||
t.Fatalf("evaluation=%+v err=%v", value, err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestBuildAdditionalProfileRequiresUnresolvedEvidenceAndPreservesHistory(t *testing.T) {
|
||||
parent := capability.TaskEvaluationProfile{ID: "general-v1", Revision: 1, CapabilityID: "creator-advisory", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"general"}, QualityWeightMillis: 700, ReliabilityWeightMillis: 200, LatencyWeightMillis: 100}
|
||||
profile, err := BuildAdditionalTaskEvaluationProfile(parent, capability.TaskEvaluationExpansionRequest{ParentProfileID: parent.ID, CapabilityID: parent.CapabilityID, TriggerEvidenceRef: "experience-unsolved-1", NewCaseIDs: []string{"edge-case"}, BenchmarkVersion: "additional-v1"}, "additional-v1", 1)
|
||||
if err != nil || profile.Stage != capability.TaskEvaluationAdditional || profile.ParentProfileID != parent.ID || profile.RequiredCaseIDs[0] != "edge-case" {
|
||||
t.Fatalf("profile=%+v err=%v", profile, err)
|
||||
}
|
||||
}
|
||||
@@ -18,6 +18,7 @@ type Repository interface {
|
||||
SaveModelDescriptor(context.Context, capability.ModelDescriptor) error
|
||||
SaveCognitiveBinding(context.Context, capability.CognitiveBinding) error
|
||||
LoadCognitiveCandidates(context.Context, capability.CapabilityID, identity.PrincipalID) ([]capability.ResolvedCognitiveBinding, error)
|
||||
LoadCognitiveCandidatesForProfile(context.Context, capability.CapabilityID, identity.PrincipalID, capability.TaskEvaluationProfileID) ([]capability.ResolvedCognitiveBinding, error)
|
||||
LoadCognitiveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error)
|
||||
SaveCognitiveCanaryPromotion(context.Context, capability.CognitiveCanaryPromotion) error
|
||||
SaveCognitiveActivationPromotion(context.Context, capability.CognitiveActivationPromotion) error
|
||||
@@ -25,4 +26,5 @@ type Repository interface {
|
||||
SaveCognitiveDiscovery(context.Context, capability.CognitiveDiscoveryRegistration) error
|
||||
SaveEvaluatedCapability(context.Context, capability.EvaluatedCapability) error
|
||||
SavePerformanceEvidence(context.Context, capability.PerformanceEvidence) error
|
||||
SaveTaskEvaluation(context.Context, capability.TaskEvaluation) error
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user