cmd/agent-worker/task_evaluation_executor.go

@@ -50,6 +50,17 @@ func (e taskEvaluationExecutor) Execute(ctx context.Context, request work.AgentE
 	if len(response.Responses) == 0 || response.ObservedAtUnixMillis < 1 {
 		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation returned no observations")
 	}
+	if len(response.Observations) == len(intent.TaskEvaluationCases) && taskEvaluationHasCriticalFailure(response.Observations) {
+		evaluation, err := capabilitydomain.BuildTaskEvaluation(*intent.TaskEvaluationProfile, capability.TaskEvaluationID("evaluation:"+intent.DispatchID), response.BindingID, response.BindingRevision, intent.TaskEvaluationProfile.CapabilityID, response.Observations, artifact.ArtifactID(intent.TaskEvaluationCases[0].EvidenceArtifactRef), response.ObservedAtUnixMillis)
+		if err != nil {
+			return work.AgentExecutionResult{}, err
+		}
+		callback := work.TaskEvaluationEvidenceCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, NodeID: execution.NodeID, Evaluation: evaluation, Succeeded: true}
+		if _, err := postJSON[json.RawMessage](ctx, e.client, e.callbackURL+"/v1/work-graphs/task-evaluations/callback", callback); err != nil {
+			return work.AgentExecutionResult{}, err
+		}
+		return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "task evaluation recorded as UNKNOWN after bounded provider recovery", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+	}
 	if len(intent.TaskEvaluationCases) == 0 {
 		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation cases are missing")
 	}
@@ -80,6 +91,15 @@ func (e taskEvaluationExecutor) Execute(ctx context.Context, request work.AgentE
 	return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "task evaluation executed, aggregated and accepted by Work Graph callback", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
 }
 
+func taskEvaluationHasCriticalFailure(observations []capability.TaskCaseObservation) bool {
+	for _, observation := range observations {
+		if observation.CriticalFailure {
+			return true
+		}
+	}
+	return false
+}
+
 func postJSON[T any](ctx context.Context, client *http.Client, url string, payload any) (T, error) {
 	var result T
 	body, err := json.Marshal(payload)

cmd/agent-worker/task_evaluation_executor_test.go

@@ -41,3 +41,32 @@ func TestTaskEvaluationExecutorInvokesAggregatesAndCallbacks(t *testing.T) {
 		t.Fatalf("result=%+v err=%v", result, err)
 	}
 }
+
+func TestTaskEvaluationExecutorPersistsUnknownAfterProviderFailure(t *testing.T) {
+	profile := capability.TaskEvaluationProfile{ID: "profile-unknown", Revision: 1, CapabilityID: "capability-1", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "unknown-v1", RequiredCaseIDs: []string{"case-1"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/v1/task-evaluations/execute" {
+			_ = json.NewEncoder(w).Encode(capability.TaskEvaluationExecutionResponse{ProfileID: profile.ID, BindingID: "binding-1", BindingRevision: 3, Observations: []capability.TaskCaseObservation{{CaseID: "case-1", Result: capability.EvaluationUnknown, CriticalFailure: true, EvidenceArtifactRef: "artifact-1"}}, Responses: []capability.TaskCaseResponse{{CaseID: "case-1", EvidenceArtifactRef: "artifact-1"}}, ObservedAtUnixMillis: 10})
+			return
+		}
+		if r.URL.Path == "/v1/work-graphs/task-evaluations/callback" {
+			var callback work.TaskEvaluationEvidenceCallback
+			if err := json.NewDecoder(r.Body).Decode(&callback); err != nil || callback.Evaluation.Result != capability.EvaluationUnknown || !callback.Succeeded {
+				http.Error(w, "invalid unknown callback", http.StatusBadRequest)
+				return
+			}
+			w.WriteHeader(http.StatusNoContent)
+			return
+		}
+		if r.URL.Path == "/v1/task-evaluations/verify" {
+			t.Errorf("verifier must not be called for a provider failure")
+		}
+		http.NotFound(w, r)
+	}))
+	defer server.Close()
+	executor := newTaskEvaluationExecutor(server.URL, server.URL, server.URL, server.Client())
+	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-unknown", TaskEvaluationBindingID: "binding-1", TaskEvaluationProfile: &profile, TaskEvaluationCases: []capability.TaskEvaluationCase{{CaseID: "case-1", Prompt: "answer", EvidenceArtifactRef: "artifact-1"}}, TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 1, NodeID: "node-1", BindingRevision: 3, ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1"}}})
+	if err != nil || result.State != work.AgentExecutionCompleted || result.OutcomeRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}

cmd/cognitive-worker/task_evaluation.go

@@ -1,9 +1,11 @@
 package main
 
 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
+	"strings"
 	"sync"
 	"time"
 
@@ -34,20 +36,13 @@ func (g cognitiveGateway) taskEvaluationHandler(w http.ResponseWriter, r *http.R
 		http.Error(w, "external egress acknowledgement required", http.StatusBadRequest)
 		return
 	}
-	provider := g.provider(selected)
-	if provider == nil {
-		http.Error(w, "cognitive provider unavailable", http.StatusServiceUnavailable)
-		return
-	}
+
 	caseSet := make(map[string]struct{}, len(request.Profile.RequiredCaseIDs))
+	seenCases := make(map[string]struct{}, len(request.Cases))
 	for _, id := range request.Profile.RequiredCaseIDs {
 		caseSet[id] = struct{}{}
 	}
-	observations := make([]capability.TaskCaseObservation, len(request.Cases))
-	responses := make([]capability.TaskCaseResponse, len(request.Cases))
-	var wg sync.WaitGroup
-	seenCases := make(map[string]struct{}, len(request.Cases))
-	for i, item := range request.Cases {
+	for _, item := range request.Cases {
 		if item.CaseID == "" || item.Prompt == "" || item.EvidenceArtifactRef == "" {
 			http.Error(w, "invalid task evaluation case", http.StatusBadRequest)
 			return
@@ -61,35 +56,83 @@ func (g cognitiveGateway) taskEvaluationHandler(w http.ResponseWriter, r *http.R
 			return
 		}
 		seenCases[item.CaseID] = struct{}{}
-		wg.Add(1)
-		go func(index int, item capability.TaskEvaluationCase) {
-			defer wg.Done()
-			started := time.Now()
-			invocation := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("task-evaluation:%s:%s", request.Profile.ID, item.CaseID)), ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: item.Prompt, MaxOutputBytes: int(request.MaxOutputBytes), ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
-			response, invokeErr := g.invokeAdmitted(r.Context(), string(invocation.RequestID), provider, invocation, int(request.MaxOutputBytes))
-			observation := capability.TaskCaseObservation{CaseID: item.CaseID, LatencyMicros: time.Since(started).Microseconds(), EvidenceArtifactRef: item.EvidenceArtifactRef}
-			responses[index] = capability.TaskCaseResponse{CaseID: item.CaseID, ResponseText: response.AdvisoryText, LatencyMicros: observation.LatencyMicros, EvidenceArtifactRef: item.EvidenceArtifactRef}
-			if invokeErr == nil && response.FinishReason == cognitive.FinishCompleted && response.AdvisoryText != "" {
-				// Invocation success proves reachability only. Without an
-				// independent verifier and answer evidence, semantic support and
-				// quality must remain UNKNOWN rather than becoming false evidence.
-				observation.Result = capability.EvaluationUnknown
-				observation.QualityScoreMillis = 0
-			} else {
-				observation.Result = capability.EvaluationUnknown
-				observation.CriticalFailure = invokeErr != nil
-			}
-			observations[index] = observation
-		}(i, item)
 	}
 	if len(seenCases) != len(caseSet) {
 		http.Error(w, "task evaluation case set is incomplete", http.StatusBadRequest)
 		return
 	}
-	wg.Wait()
+
+	// One bounded fan-out wave uses one binding snapshot. If that candidate is
+	// unavailable, only this binding is excluded and the existing resolver
+	// chooses the next eligible candidate. Provider output remains untrusted
+	// until the independent verifier accepts it downstream.
+	requirement := capability.CognitiveRequirement{CapabilityID: request.CapabilityID, OwnerPrincipalID: request.PrincipalID, DataPolicy: selected.Binding.DataPolicy, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, AllowCanary: g.allowCanary, MaxOutputBytes: request.MaxOutputBytes}
+	var observations []capability.TaskCaseObservation
+	var responses []capability.TaskCaseResponse
+	for attempt := 0; attempt < 3; attempt++ {
+		provider := g.provider(selected)
+		if provider == nil {
+			observations, responses = taskEvaluationUnavailable(request.Cases)
+		} else {
+			observations, responses = g.executeTaskEvaluationCases(r.Context(), request, selected, provider)
+		}
+		if taskEvaluationResponsesUsable(responses) {
+			break
+		}
+		requirement.ExcludedBindingIDs = append(requirement.ExcludedBindingIDs, selected.Binding.ID)
+		fallback, resolveErr := g.resolver.Resolve(r.Context(), requirement)
+		fallbackProvider := g.provider(fallback)
+		if resolveErr != nil || fallbackProvider == nil || fallback.Binding.ID == selected.Binding.ID || (fallback.Binding.DataPolicy != capability.DataPolicyLocalOnly && !request.EgressAcknowledged) {
+			break
+		}
+		selected = fallback
+	}
+
 	writeTaskEvaluationJSON(w, capability.TaskEvaluationExecutionResponse{ProfileID: request.Profile.ID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, Observations: observations, Responses: responses, ObservedAtUnixMillis: time.Now().UnixMilli()})
 }
 
+func (g cognitiveGateway) executeTaskEvaluationCases(ctx context.Context, request capability.TaskEvaluationExecutionRequest, selected capability.ResolvedCognitiveBinding, provider advisoryProvider) ([]capability.TaskCaseObservation, []capability.TaskCaseResponse) {
+	observations := make([]capability.TaskCaseObservation, len(request.Cases))
+	responses := make([]capability.TaskCaseResponse, len(request.Cases))
+	var wg sync.WaitGroup
+	for i, item := range request.Cases {
+		wg.Add(1)
+		go func(index int, item capability.TaskEvaluationCase) {
+			defer wg.Done()
+			started := time.Now()
+			invocation := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("task-evaluation:%s:%s", request.Profile.ID, item.CaseID)), ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: item.Prompt, MaxOutputBytes: int(request.MaxOutputBytes), ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
+			response, invokeErr := g.invokeAdmitted(ctx, string(invocation.RequestID), provider, invocation, int(request.MaxOutputBytes))
+			latency := time.Since(started).Microseconds()
+			observations[index] = capability.TaskCaseObservation{CaseID: item.CaseID, Result: capability.EvaluationUnknown, CriticalFailure: invokeErr != nil || response.FinishReason != cognitive.FinishCompleted || strings.TrimSpace(response.AdvisoryText) == "", EvidenceArtifactRef: item.EvidenceArtifactRef, LatencyMicros: latency}
+			responses[index] = capability.TaskCaseResponse{CaseID: item.CaseID, ResponseText: response.AdvisoryText, LatencyMicros: latency, EvidenceArtifactRef: item.EvidenceArtifactRef}
+		}(i, item)
+	}
+	wg.Wait()
+	return observations, responses
+}
+
+func taskEvaluationUnavailable(cases []capability.TaskEvaluationCase) ([]capability.TaskCaseObservation, []capability.TaskCaseResponse) {
+	observations := make([]capability.TaskCaseObservation, len(cases))
+	responses := make([]capability.TaskCaseResponse, len(cases))
+	for i, item := range cases {
+		observations[i] = capability.TaskCaseObservation{CaseID: item.CaseID, Result: capability.EvaluationUnknown, CriticalFailure: true, EvidenceArtifactRef: item.EvidenceArtifactRef}
+		responses[i] = capability.TaskCaseResponse{CaseID: item.CaseID, EvidenceArtifactRef: item.EvidenceArtifactRef}
+	}
+	return observations, responses
+}
+
+func taskEvaluationResponsesUsable(responses []capability.TaskCaseResponse) bool {
+	if len(responses) == 0 {
+		return false
+	}
+	for _, response := range responses {
+		if strings.TrimSpace(response.ResponseText) == "" {
+			return false
+		}
+	}
+	return true
+}
+
 func writeTaskEvaluationJSON(w http.ResponseWriter, value any) {
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(value)

cmd/cognitive-worker/task_evaluation_test.go

@@ -56,3 +56,90 @@ func TestTaskEvaluationHandlerInvokesEveryBoundedCase(t *testing.T) {
 		}
 	}
 }
+
+type taskEvaluationFailoverRegistry struct {
+	primary, fallback capability.ResolvedCognitiveBinding
+	resolveCalls      int
+}
+
+func (s *taskEvaluationFailoverRegistry) ResolveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error) {
+	return s.primary, nil
+}
+
+func (s *taskEvaluationFailoverRegistry) Resolve(_ context.Context, requirement capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error) {
+	s.resolveCalls++
+	if len(requirement.ExcludedBindingIDs) == 0 {
+		return s.primary, nil
+	}
+	return s.fallback, nil
+}
+
+func (s *taskEvaluationFailoverRegistry) Observe(context.Context, capability.CognitiveUsageObservation) error {
+	return nil
+}
+
+func (s *taskEvaluationFailoverRegistry) RegisterDiscovery(context.Context, capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error) {
+	return capability.CognitiveDiscoveryRegistration{}, nil
+}
+
+func TestTaskEvaluationExcludesUnavailableCandidateAndUsesFallback(t *testing.T) {
+	primary := gatewaySelection()
+	fallback := primary
+	fallback.Binding.ID = "binding-fallback"
+	fallback.Binding.Revision = 7
+	fallback.Binding.ProviderInstanceID = "provider-fallback"
+	fallback.Provider.ID = "provider-fallback"
+	fallback.Provider.Kind = "provider-fallback-kind"
+	fallback.Provider.EndpointRef = "fallback-endpoint"
+	fallback.Provider.CredentialRef = "fallback-credential"
+	fallback.Model.ProviderInstanceID = "provider-fallback"
+	registry := &taskEvaluationFailoverRegistry{primary: primary, fallback: fallback}
+	primaryProvider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishProviderUnavailable}}
+	fallbackProvider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishCompleted, AdvisoryText: "fallback response"}}
+	gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{
+		primary.Provider.Kind:  {{ProviderInstanceID: primary.Provider.ID, Kind: primary.Provider.Kind, EndpointRef: primary.Provider.EndpointRef, CredentialRef: primary.Provider.CredentialRef, Provider: primaryProvider}},
+		fallback.Provider.Kind: {{ProviderInstanceID: fallback.Provider.ID, Kind: fallback.Provider.Kind, EndpointRef: fallback.Provider.EndpointRef, CredentialRef: fallback.Provider.CredentialRef, Provider: fallbackProvider}},
+	}, now: func() time.Time { return time.UnixMilli(1000) }}
+	profile := capability.TaskEvaluationProfile{ID: "profile-failover", Revision: 1, CapabilityID: primary.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "failover-v1", RequiredCaseIDs: []string{"case-a"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: profile.CapabilityID, BindingID: primary.Binding.ID, BindingRevision: primary.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: primary.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-a", Prompt: "one", EvidenceArtifactRef: "artifact:a"}}}
+	payload, _ := json.Marshal(request)
+	recorder := httptest.NewRecorder()
+	gateway.taskEvaluationHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/execute", bytes.NewReader(payload)))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+	var response capability.TaskEvaluationExecutionResponse
+	if err := json.Unmarshal(recorder.Body.Bytes(), &response); err != nil {
+		t.Fatal(err)
+	}
+	if response.BindingID != fallback.Binding.ID || !primaryProvider.called || !fallbackProvider.called || registry.resolveCalls != 1 {
+		t.Fatalf("failover binding=%s primary_called=%t fallback_called=%t resolve_calls=%d", response.BindingID, primaryProvider.called, fallbackProvider.called, registry.resolveCalls)
+	}
+	if response.Observations[0].CriticalFailure || response.Responses[0].ResponseText != "fallback response" {
+		t.Fatalf("fallback response was not usable: %+v %+v", response.Observations, response.Responses)
+	}
+}
+
+func TestTaskEvaluationUnavailableResponseIsMarkedCriticalUnknown(t *testing.T) {
+	selected := gatewaySelection()
+	registry := &taskEvaluationFailoverRegistry{primary: selected, fallback: selected}
+	provider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishTimeout}}
+	gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{
+		selected.Provider.Kind: {{ProviderInstanceID: selected.Provider.ID, Kind: selected.Provider.Kind, EndpointRef: selected.Provider.EndpointRef, CredentialRef: selected.Provider.CredentialRef, Provider: provider}},
+	}, now: func() time.Time { return time.UnixMilli(1000) }}
+	profile := capability.TaskEvaluationProfile{ID: "profile-unknown", Revision: 1, CapabilityID: selected.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "unknown-v1", RequiredCaseIDs: []string{"case-a"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: profile.CapabilityID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: selected.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-a", Prompt: "one", EvidenceArtifactRef: "artifact:a"}}}
+	payload, _ := json.Marshal(request)
+	recorder := httptest.NewRecorder()
+	gateway.taskEvaluationHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/execute", bytes.NewReader(payload)))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+	var response capability.TaskEvaluationExecutionResponse
+	if err := json.Unmarshal(recorder.Body.Bytes(), &response); err != nil {
+		t.Fatal(err)
+	}
+	if len(response.Observations) != 1 || response.Observations[0].Result != capability.EvaluationUnknown || !response.Observations[0].CriticalFailure {
+		t.Fatalf("unavailable response was overclaimed: %+v", response.Observations)
+	}
+}

dev/ACTIVE_TASK.json

@@ -1,5 +1,5 @@
 {
-  "task_file": "dev/tasks/DEV-000248-live-task-evaluation-scheduling.json",
-  "task_id": "DEV-000248",
+  "task_file": "dev/tasks/DEV-000249-task-evaluation-model-failover.json",
+  "task_id": "DEV-000249",
   "task_revision": 1
 }

dev/tasks/DEV-000249-task-evaluation-model-failover.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000249",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "f90ed7af35f9f19c72039812272640ae1d63635b",
+  "risk_tier": "E",
+  "intent": "Make task-evaluation execution use the existing cognitive model resolution and local fallback path without recording unsupported output as qualification evidence.",
+  "requirement_ids": ["model fleet/failover", "live task evaluation", "local recovery"],
+  "allowed_change_scope": {"modules": ["capability", "work"], "paths": ["cmd/cognitive-worker/**", "cmd/agent-worker/**", "modules/capability/domain/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000249-task-evaluation-model-failover.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "LLM-local/**", "modules/work/**", "modules/work/secrets/**"]},
+  "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": ["existing CognitiveRequirement", "existing ResolvedCognitiveBinding", "existing TaskEvaluationExecutionRequest/Response"], "may_change": []},
+  "semantic_owner": "capability cognitive execution adapter",
+  "canonical_state": ["existing cognitive usage observations and task evaluation records only"],
+  "state_transitions": ["candidate invocation -> bounded fallback candidate -> verified evaluation or UNKNOWN"],
+  "side_effects_and_privileges": ["existing cognitive provider calls and capability observations; no new privileges"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["provider failure excludes the failed binding for the current bounded execution; all candidates unavailable returns UNKNOWN without qualification"],
+  "assumptions": ["fallback resolution remains governed by the existing capability resolver and data-policy checks"],
+  "acceptance_scenarios": ["task evaluation invokes active candidate", "provider-unavailable candidate is excluded and local candidate is attempted", "empty/non-completed provider response never becomes supported evidence", "all candidates unavailable produces durable UNKNOWN evaluation and terminal graph callback"],
+  "required_tests": ["go test ./cmd/cognitive-worker ./modules/capability/domain", "go vet ./cmd/cognitive-worker ./modules/capability/domain", "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", "live task-evaluation canary"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}