Enforce classifier-error fail-closed policy

README.md

@@ -141,6 +141,14 @@ Adrian supports entirely offline, data sovereign deployments using just a handfu
 
    Use the same `adrian.init` snippet as in the [Quickstart](#quickstart) above. The SDK defaults to `ws://localhost:8080/ws`, so a self-hosted setup needs nothing more than the API key - drop the `ws_url=` line.
 
+### Classifier error policy
+
+Adrian records classifier outages, malformed classifier responses, and unparseable classifier output as `verdict_status=error` with `mad_code=""`. These are operational classifier errors, not benign `M0` findings and not synthetic malicious activity.
+
+The default policy remains availability-first: classifier errors fail open. In **Settings -> Policy**, enable **Fail closed on classifier error** to make BLOCK-mode tool calls return blocked responses when the classifier cannot produce a verdict. In HITL mode, actionable classifier errors are sent to the review queue and held until an operator approves or rejects them.
+
+Fail-closed classifier-error enforcement requires the Python SDK version shipped with this repository update. Older SDKs ignore the additive protobuf `status` and policy fields, see an empty MAD code, and continue fail-open even when the dashboard toggle is enabled.
+
 To [reset the admin password](https://docs.adrian.secureagentics.ai/reference/backend#reset-the-admin-password), [change the model](https://docs.adrian.secureagentics.ai/reference/backend#switch-the-local-gguf) and much more check out the dedicated [Docs site](https://docs.adrian.secureagentics.ai/).
 
 ## Why Adrian is different

backend/internal/api/handlers_events.go

@@ -84,13 +84,18 @@ func (s *Server) handleListEvents(w http.ResponseWriter, r *http.Request) {
 			since = t
 		}
 	}
+	if status := q.Get("verdict_status"); status != "" && !validVerdictStatus(status) {
+		writeError(w, http.StatusBadRequest, "invalid verdict_status")
+		return
+	}
 
 	filters := store.EventFilters{
-		Since:     since,
-		AgentID:   q.Get("agent_id"),
-		SessionID: q.Get("session_id"),
-		EventType: q.Get("event_type"),
-		MinMAD:    q.Get("min_mad"),
+		Since:         since,
+		AgentID:       q.Get("agent_id"),
+		SessionID:     q.Get("session_id"),
+		EventType:     q.Get("event_type"),
+		MinMAD:        q.Get("min_mad"),
+		VerdictStatus: q.Get("verdict_status"),
 	}
 
 	rows, total, err := s.store.ListEvents(r.Context(), filters, pg.PerPage, pg.Offset)
@@ -203,6 +208,7 @@ func eventToListItemResponse(r *store.EventListRow) eventListItemResponse {
 			ID:             r.VerdictID,
 			MADCode:        r.MADCode,
 			Classification: r.Classification,
+			VerdictStatus:  r.VerdictStatus,
 		}
 	}
 	return resp

backend/internal/api/handlers_reviews.go

@@ -39,6 +39,7 @@ type reviewDetail struct {
 	reviewSummary
 	EventPayload   json.RawMessage `json:"event_payload,omitempty"`
 	Classification string          `json:"classification,omitempty"`
+	Reasoning      string          `json:"reasoning,omitempty"`
 }
 
 type reviewResolveResponse struct {
@@ -49,8 +50,14 @@ type reviewResolveResponse struct {
 
 func (s *Server) handleListReviews(w http.ResponseWriter, r *http.Request) {
 	pg := parsePagination(r)
-	status := r.URL.Query().Get("status")
-	rows, total, err := s.store.ListHitlQueue(r.Context(), status, pg.PerPage, pg.Offset)
+	q := r.URL.Query()
+	status := q.Get("status")
+	verdictStatus := q.Get("verdict_status")
+	if verdictStatus != "" && !validVerdictStatus(verdictStatus) {
+		writeError(w, http.StatusBadRequest, "invalid verdict_status")
+		return
+	}
+	rows, total, err := s.store.ListHitlQueue(r.Context(), status, verdictStatus, pg.PerPage, pg.Offset)
 	if err != nil {
 		writeError(w, http.StatusInternalServerError, "query failed")
 		return
@@ -81,6 +88,7 @@ func (s *Server) handleGetReview(w http.ResponseWriter, r *http.Request) {
 	resp := reviewDetail{
 		reviewSummary:  reviewToSummary(&row.HitlReview),
 		Classification: row.Classification,
+		Reasoning:      row.Reasoning,
 	}
 	if row.EventPayloadJSON != "" {
 		resp.EventPayload = json.RawMessage(row.EventPayloadJSON)
@@ -128,7 +136,7 @@ func (s *Server) resolveReview(w http.ResponseWriter, r *http.Request, status st
 			EventId:   row.EventID,
 			SessionId: row.SessionID,
 			MadCode:   row.MADCode,
-			Status:    pb.VerdictStatus_VERDICT_STATUS_OK,
+			Status:    reviewVerdictStatusProto(row.VerdictStatus),
 			Policy:    s.policySnapshotProto(pol),
 			Hitl:      &pb.HitlResponse{ContinueExecution: continueExec},
 		}},
@@ -165,3 +173,14 @@ func reviewToSummary(r *store.HitlReview) reviewSummary {
 	}
 	return out
 }
+
+func reviewVerdictStatusProto(status string) pb.VerdictStatus {
+	switch status {
+	case "error":
+		return pb.VerdictStatus_VERDICT_STATUS_ERROR
+	case "ok":
+		return pb.VerdictStatus_VERDICT_STATUS_OK
+	default:
+		return pb.VerdictStatus_VERDICT_STATUS_UNSPECIFIED
+	}
+}

backend/internal/api/handlers_stats.go

@@ -6,12 +6,13 @@ package api
 import "net/http"
 
 type overviewResponse struct {
-	TotalEvents     int            `json:"total_events"`
-	FlaggedVerdicts int            `json:"flagged_verdicts"`
-	PendingReviews  int            `json:"pending_reviews"`
-	ActiveAgents    int            `json:"active_agents"`
-	VerdictsByMAD   map[string]int `json:"verdicts_by_mad"`
-	Window          string         `json:"window"`
+	TotalEvents      int            `json:"total_events"`
+	FlaggedVerdicts  int            `json:"flagged_verdicts"`
+	ClassifierErrors int            `json:"classifier_errors"`
+	PendingReviews   int            `json:"pending_reviews"`
+	ActiveAgents     int            `json:"active_agents"`
+	VerdictsByMAD    map[string]int `json:"verdicts_by_mad"`
+	Window           string         `json:"window"`
 }
 
 type activityBucketEntry struct {
@@ -31,12 +32,13 @@ func (s *Server) handleStatsOverview(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	writeJSON(w, http.StatusOK, overviewResponse{
-		TotalEvents:     o.TotalEvents,
-		FlaggedVerdicts: o.FlaggedVerdicts,
-		PendingReviews:  o.PendingReviews,
-		ActiveAgents:    o.ActiveAgents,
-		VerdictsByMAD:   o.VerdictsByMAD,
-		Window:          "24h",
+		TotalEvents:      o.TotalEvents,
+		FlaggedVerdicts:  o.FlaggedVerdicts,
+		ClassifierErrors: o.ClassifierErrors,
+		PendingReviews:   o.PendingReviews,
+		ActiveAgents:     o.ActiveAgents,
+		VerdictsByMAD:    o.VerdictsByMAD,
+		Window:           "24h",
 	})
 }