feat(capability): consume qualification evidence in selection #220
@@ -505,6 +505,14 @@ func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value
|
||||
if err == nil {
|
||||
value.Performance = &performance
|
||||
}
|
||||
var qualification capability.ModelQualificationObservation
|
||||
err = r.pool.QueryRow(ctx, `SELECT observation_id,provider_instance_id,provider_model_id,status,finish_reason,latency_millis,output_observed,quality_assessed,category,observed_at_unix_millis FROM capability.cognitive_model_qualification_observations WHERE provider_instance_id=$1 AND provider_model_id=$2 ORDER BY observed_at_unix_millis DESC LIMIT 1`, value.Provider.ID, value.Model.ProviderModelID).Scan(&qualification.ID, &qualification.ProviderInstanceID, &qualification.ProviderModelID, &qualification.Status, &qualification.FinishReason, &qualification.LatencyMillis, &qualification.OutputObserved, &qualification.QualityAssessed, &qualification.Category, &qualification.ObservedAtUnixMillis)
|
||||
if err != nil && !errors.Is(err, pgx.ErrNoRows) {
|
||||
return err
|
||||
}
|
||||
if err == nil {
|
||||
value.Qualification = &qualification
|
||||
}
|
||||
if profileID != "" {
|
||||
var taskEvaluation capability.TaskEvaluation
|
||||
err = r.pool.QueryRow(ctx, `SELECT evaluation_id,task_profile_id,binding_id,binding_revision,capability_id,result,quality_score_millis,reliability_score_millis,latency_score_millis,overall_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis FROM capability.cognitive_task_evaluations WHERE task_profile_id=$1 AND binding_id=$2 AND binding_revision=$3 ORDER BY observed_at_unix_millis DESC LIMIT 1`, profileID, value.Binding.ID, value.Binding.Revision).Scan(&taskEvaluation.ID, &taskEvaluation.ProfileID, &taskEvaluation.BindingID, &taskEvaluation.BindingRevision, &taskEvaluation.CapabilityID, &taskEvaluation.Result, &taskEvaluation.QualityScoreMillis, &taskEvaluation.ReliabilityScoreMillis, &taskEvaluation.LatencyScoreMillis, &taskEvaluation.OverallScoreMillis, &taskEvaluation.CriticalFailureMillis, &taskEvaluation.BenchmarkVersion, &taskEvaluation.ArtifactID, &taskEvaluation.ObservedAtUnixMillis)
|
||||
|
||||
@@ -172,6 +172,7 @@ type ResolvedCognitiveBinding struct {
|
||||
Evaluation *EvaluatedCapability
|
||||
Performance *PerformanceEvidence
|
||||
TaskEvaluation *TaskEvaluation
|
||||
Qualification *ModelQualificationObservation
|
||||
Rating *ModelRating
|
||||
}
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
{
|
||||
"task_file": "dev/tasks/DEV-000242-durable-qualification-observations.json",
|
||||
"task_id": "DEV-000242",
|
||||
"task_file": "dev/tasks/DEV-000243-qualification-selection-evidence.json",
|
||||
"task_id": "DEV-000243",
|
||||
"task_revision": 1
|
||||
}
|
||||
|
||||
@@ -0,0 +1,23 @@
|
||||
{
|
||||
"task_id": "DEV-000243",
|
||||
"task_revision": 1,
|
||||
"architecture_revision": "ADR-0018",
|
||||
"base_repo_revision": "47327cc6d853da9084c96e3bf48502266706ca57",
|
||||
"risk_tier": "E",
|
||||
"intent": "Make Capability selection consume the latest durable qualification observation for fresh unavailability exclusion and transparent latency rating, without granting quality or activation.",
|
||||
"requirement_ids": ["qualification selection loop", "hardware-neutral routing", "quality separated from reachability"],
|
||||
"allowed_change_scope": {"modules": ["capability", "adapters/postgres"], "paths": ["adapters/postgres/capability_repository.go", "contracts/capability/cognitive_registry.go", "modules/capability/domain/cognitive_registry.go", "modules/capability/domain/model_rating_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000243-qualification-selection-evidence.json"]},
|
||||
"forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/work/**"]},
|
||||
"allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
|
||||
"contracts": {"consume": ["durable ModelQualificationObservation"], "may_change": ["resolved binding evidence projection"]},
|
||||
"semantic_owner": "Capability selection and rating",
|
||||
"canonical_state": ["qualification observation remains evidence; no activation mutation"],
|
||||
"state_transitions": ["fresh UNAVAILABLE observation -> candidate excluded", "stale UNAVAILABLE observation -> discovery state remains authoritative", "REACHABLE observation -> latency evidence only"],
|
||||
"side_effects_and_privileges": ["read-only projection and deterministic selection"],
|
||||
"persistence_and_migrations": ["uses migration 045 from DEV-000242"],
|
||||
"failure_retry_restart": ["missing observation preserves existing behavior"],
|
||||
"assumptions": ["provider/model observed timestamps define freshness boundary"],
|
||||
"acceptance_scenarios": ["fresh unavailable candidate is not selected", "stale unavailable candidate is not permanently poisoned", "quality remains task-evaluation controlled"],
|
||||
"required_tests": ["go test ./modules/capability/... ./adapters/postgres/...", "go vet ./modules/capability/... ./adapters/postgres/...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture", "git diff --check", "independent", "CI", "production health smoke"],
|
||||
"terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
|
||||
}
|
||||
@@ -148,6 +148,9 @@ func SelectCognitiveBinding(requirement capability.CognitiveRequirement, candida
|
||||
if candidate.Evaluation != nil && candidate.Evaluation.Result == capability.EvaluationUnsupported {
|
||||
continue
|
||||
}
|
||||
if candidate.Qualification != nil && candidate.Qualification.Status == capability.ModelQualificationUnavailable && candidate.Qualification.ObservedAtUnixMillis >= maxObservedAt(candidate.Provider.ObservedAtUnixMillis, candidate.Model.ObservedAtUnixMillis) {
|
||||
continue
|
||||
}
|
||||
if requirement.TaskProfileID != "" && (candidate.TaskEvaluation == nil || candidate.TaskEvaluation.ProfileID != requirement.TaskProfileID || candidate.TaskEvaluation.Result != capability.EvaluationSupported) {
|
||||
continue
|
||||
}
|
||||
@@ -195,10 +198,14 @@ func RateCognitiveBinding(candidate capability.ResolvedCognitiveBinding) capabil
|
||||
latency := int64(500)
|
||||
if candidate.Performance != nil && candidate.Performance.LatencyMicros != nil {
|
||||
latency = latencyScore(*candidate.Performance.LatencyMicros)
|
||||
} else if candidate.Qualification != nil && candidate.Qualification.Status == capability.ModelQualificationReachable {
|
||||
latency = latencyScore(candidate.Qualification.LatencyMillis * 1000)
|
||||
}
|
||||
evidence := int64(700)
|
||||
if candidate.Performance != nil {
|
||||
evidence = 1000
|
||||
} else if candidate.Qualification != nil && candidate.Qualification.Status == capability.ModelQualificationReachable {
|
||||
evidence = 750
|
||||
}
|
||||
overall := (quality*55 + reliability*25 + latency*15 + evidence*5) / 100
|
||||
category := capability.ModelRatingBalanced
|
||||
@@ -210,6 +217,13 @@ func RateCognitiveBinding(candidate capability.ResolvedCognitiveBinding) capabil
|
||||
return capability.ModelRating{Category: category, OverallScoreMillis: overall, QualityScoreMillis: quality, ReliabilityScoreMillis: reliability, LatencyScoreMillis: latency, EvidenceScoreMillis: evidence}
|
||||
}
|
||||
|
||||
func maxObservedAt(left, right int64) int64 {
|
||||
if left > right {
|
||||
return left
|
||||
}
|
||||
return right
|
||||
}
|
||||
|
||||
func clampMillis(value int64) int64 {
|
||||
if value < 0 {
|
||||
return 0
|
||||
|
||||
@@ -38,3 +38,22 @@ func TestRateCognitiveBindingClassifiesSlowEvidenceAsBalanced(t *testing.T) {
|
||||
t.Fatalf("rating=%+v", rating)
|
||||
}
|
||||
}
|
||||
|
||||
func TestSelectionExcludesFreshUnavailableQualificationButNotStaleEvidence(t *testing.T) {
|
||||
quality, critical := int64(900), int64(0)
|
||||
base := capability.ResolvedCognitiveBinding{
|
||||
Binding: capability.CognitiveBinding{ID: "binding-qualification", Revision: 1, CapabilityID: "creator-advisory", OwnerPrincipalID: "owner", ProviderInstanceID: "provider", ProviderRevision: 1, ModelDescriptorID: "model", ModelRevision: 1, DataPolicy: capability.DataPolicyLocalOnly, Qualification: capability.QualificationValidated, Activation: capability.ActivationActive, Health: capability.HealthAvailable, LimitPolicyRef: "limits"},
|
||||
Provider: capability.ProviderInstance{ID: "provider", Revision: 1, Kind: "local", EndpointRef: "endpoint", Enabled: true, Qualification: capability.QualificationValidated, Health: capability.HealthAvailable, ObservedAtUnixMillis: 20},
|
||||
Model: capability.ModelDescriptor{ID: "model", Revision: 1, ProviderInstanceID: "provider", ProviderModelID: "model", MaxContextBytes: 1000, Modalities: []capability.ModelModality{capability.ModalityText}, Capabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, Qualification: capability.QualificationValidated, Health: capability.HealthAvailable, ObservedAtUnixMillis: 20},
|
||||
Evaluation: &capability.EvaluatedCapability{ID: "evaluation", BindingID: "binding-qualification", BindingRevision: 1, CapabilityID: "creator-advisory", Result: capability.EvaluationSupported, QualityScoreMillis: &quality, CriticalFailureMillis: &critical},
|
||||
Qualification: &capability.ModelQualificationObservation{ID: "observation", ProviderInstanceID: "provider", ProviderModelID: "model", Status: capability.ModelQualificationUnavailable, FinishReason: "TIMEOUT", ObservedAtUnixMillis: 20, Category: capability.ModelRatingUnrated},
|
||||
}
|
||||
requirement := capability.CognitiveRequirement{CapabilityID: "creator-advisory", OwnerPrincipalID: "owner", DataPolicy: capability.DataPolicyLocalOnly, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, MaxOutputBytes: 10}
|
||||
if _, err := SelectCognitiveBinding(requirement, []capability.ResolvedCognitiveBinding{base}); err == nil {
|
||||
t.Fatal("fresh unavailable qualification should exclude candidate")
|
||||
}
|
||||
base.Qualification.ObservedAtUnixMillis = 19
|
||||
if _, err := SelectCognitiveBinding(requirement, []capability.ResolvedCognitiveBinding{base}); err != nil {
|
||||
t.Fatalf("stale unavailable qualification should not exclude candidate: %v", err)
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user