Add adaptive learning and model rollback

Add safety dashboard and decision transparency
Fix dashboard text overflow and close v1 docs gaps
2026-06-17 18:41:03 +02:00 · 2026-06-17 18:26:49 +02:00 · 2026-06-17 11:53:25 +02:00
15 changed files with 1315 additions and 35 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,41 @@
 # Changelog

+## 1.2.0 - 2026-06-17
+- Automatische Sensor-Gewichtungsanpassung aus Nutzerfeedback:
+  korrektes Feedback staerkt aktuelle Kontextsignale leicht, falsches Feedback
+  wertet sie vorsichtig ab.
+- Modell-Snapshots mit aktivem Modellstand und Rollback-API ergaenzt.
+- Dashboard zeigt Modell-Snapshots, Rollback, Zeitprofile,
+  adaptive Gewichtungsupdates und Automation-Konflikte.
+- Automation-Refresh markiert Konflikte, wenn SillyHome aktiv ist und passende
+  HA-Automationen parallel aktiv bleiben.
+- Zeitprofile fuer Nacht, Morgen, Tag, Abend und Wochenende werden aus
+  gelernten Handlungen gebildet.
+
+## 1.1.0 - 2026-06-17
+- Dashboard als Einrichtungs- und Visualisierungszentrale erweitert:
+  Job-Queue, Sicherheitsprofil, Entscheidungsakte, Wissen/Annahmen/
+  Unsicherheiten und Beitragsfaktoren pro Aktor.
+- Lokales Safety-Profil pro Aktor eingefuehrt: manuelle Sperre,
+  Freigabestufe, Mindest-Confidence und optionaler Cooldown werden vor
+  autonomem Schalten ausgewertet.
+- Sofort-Schaltpfad bleibt schnell: Safety prueft nur lokale Daten; der
+  Home-Assistant-Serviceaufruf wird nicht durch Discovery, Training oder
+  Statistik blockiert.
+- Sichtbare Job-Queue fuer Discovery, Reconciliation, Training, Evaluation
+  und Automation-Refresh mit Status, Dauer, Fehler und Zusammenfassung.
+- Entscheidungsstatistik erweitert: Sensor-/Kontextfaktoren, aktive
+  Gewichtungen, Sample-/Confidence-Trends und Feedbackzaehler werden
+  persistiert.
+
+## 1.0.5 - 2026-06-17
+- Lange Friendly Names, Entity-IDs, Chips, Tabellenwerte und Metriken brechen
+  im Dashboard responsiv um und laufen nicht mehr aus Karten oder Boxen.
+- Automatisierter Performance-Budget-Test fuer Root-HTML und
+  `/v1/actuators/dashboard` gegen das 5-Sekunden-Limit ergaenzt.
+- HA-/Ingress-Verifikation mit Supervisor-Status, Backup, Watchdog,
+  Hard-Reload und Rollback im Operating Guide dokumentiert.
+
 ## 1.0.4 - 2026-06-17
 - Sensor-Relevanz ist in der Aktor-Detailansicht sichtbar: automatische
  Relevanz, aktive Gewichtung und Score werden pro verwendetem Sensor/Zustand
--- a/README.md
+++ b/README.md
@@ -15,6 +15,10 @@ nach einer ausdrücklichen Freigabe ausführen.
  [`docs/V1_0_0_OPERATING_GUIDE.md`](docs/V1_0_0_OPERATING_GUIDE.md)
 - Version 1.0.x Abnahme und offene Punkte:
  [`docs/V1_0_ACCEPTANCE.md`](docs/V1_0_ACCEPTANCE.md)
+- Version 1.1.0 Safety, Transparenz und Job-Queue:
+  [`docs/V1_1_0_OPERATING_GUIDE.md`](docs/V1_1_0_OPERATING_GUIDE.md)
+- Version 1.2.0 adaptive Gewichtung, Rollback und Profile:
+  [`docs/V1_2_0_OPERATING_GUIDE.md`](docs/V1_2_0_OPERATING_GUIDE.md)
 - Arbeitsregeln für Coding-Agenten: [`AGENTS.md`](AGENTS.md)

 ## Reifegrad
--- a/addon/config.yaml
+++ b/addon/config.yaml
@@ -1,5 +1,5 @@
 name: SillyHome Next
-version: "1.0.4"
+version: "1.2.0"
 slug: sillyhome_next
 description: Lernt automatisch aus deinem Verhalten und steuert freigegebene Aktoren
 url: http://192.168.6.31:3000/pino/sillyhome-next
--- a/app/actuators/models.py
+++ b/app/actuators/models.py
@@ -37,6 +37,21 @@ class BehaviorStatus(StrEnum):
    BLOCKED = "blocked"


+class SafetyStage(StrEnum):
+    OBSERVE = "observe"
+    SUGGEST = "suggest"
+    SHADOW = "shadow"
+    PARTIAL = "partial"
+    ACTIVE = "active"
+
+
+class JobStatus(StrEnum):
+    PENDING = "pending"
+    RUNNING = "running"
+    COMPLETED = "completed"
+    FAILED = "failed"
+
+
 class AssignmentCandidate(BaseModel):
    entity_id: str
    domain: str
@@ -121,11 +136,101 @@ class BehaviorPrediction(BaseModel):
    execution_reason: str = "Vorhersage wurde noch nicht ausgeführt."


+class DecisionFactor(BaseModel):
+    entity_id: str | None = None
+    label: str
+    factor_type: str = Field(max_length=40)
+    state: str | None = None
+    weight: float = Field(default=1.0, ge=0.0, le=1.0)
+    contribution: float = Field(default=0.0, ge=0.0, le=1.0)
+    evidence: list[str] = Field(default_factory=list)
+
+
+class AdaptiveWeightUpdate(BaseModel):
+    entity_id: str
+    previous_weight: float = Field(ge=0.0, le=1.0)
+    new_weight: float = Field(ge=0.0, le=1.0)
+    reason: str = Field(max_length=300)
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class SafetyRule(BaseModel):
+    rule_id: str = Field(pattern=r"^[a-z0-9_-]{1,64}$")
+    label: str = Field(min_length=1, max_length=160)
+    enabled: bool = True
+    blocking: bool = True
+    reason: str = Field(default="", max_length=300)
+
+
+def default_safety_rules() -> list[SafetyRule]:
+    return [
+        SafetyRule(
+            rule_id="activation_ready",
+            label="Nur nach Lernfreigabe aktiv schalten",
+            reason="Der Aktor muss genug eindeutiges Verhalten gelernt haben.",
+        ),
+        SafetyRule(
+            rule_id="confidence_threshold",
+            label="Mindest-Sicherheit einhalten",
+            reason="Vorhersagen unter der Schaltschwelle bleiben im Shadow-Modus.",
+        ),
+        SafetyRule(
+            rule_id="cooldown",
+            label="Sicherheits-Cooldown gegen Hin-und-her-Schalten",
+            reason="Gleiche Zielzustände werden nicht zu schnell wiederholt.",
+        ),
+        SafetyRule(
+            rule_id="manual_block",
+            label="Manuelle Sperre respektieren",
+            reason="Nutzer können jeden Aktor sofort blockieren.",
+        ),
+    ]
+
+
+class SafetyProfile(BaseModel):
+    stage: SafetyStage = SafetyStage.SHADOW
+    manual_block: bool = False
+    min_confidence: float = Field(default=0.82, ge=0.0, le=1.0)
+    min_confidence_on: float | None = Field(default=None, ge=0.0, le=1.0)
+    min_confidence_off: float | None = Field(default=None, ge=0.0, le=1.0)
+    cooldown_seconds: int | None = Field(default=None, ge=0)
+    rules: list[SafetyRule] = Field(default_factory=default_safety_rules)
+    updated_at: datetime | None = None
+    note: str | None = Field(default=None, max_length=500)
+
+
 class ExecutionEvent(BaseModel):
    target_state: str
    executed_at: datetime


+class ModelSnapshot(BaseModel):
+    version_id: str
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    sample_count: int = Field(default=0, ge=0)
+    high_confidence_sample_count: int = Field(default=0, ge=0)
+    average_confidence: float = Field(default=0.0, ge=0.0, le=1.0)
+    incorrect_feedback_count: int = Field(default=0, ge=0)
+    patterns: list[BehaviorPattern] = Field(default_factory=list)
+    reason: str = Field(default="", max_length=500)
+
+
+class AutomationConflict(BaseModel):
+    automation_entity_id: str = Field(pattern=r"^automation\.[a-z0-9_]+$")
+    severity: str = Field(default="info", max_length=20)
+    status: str = Field(default="open", max_length=40)
+    reason: str = Field(max_length=500)
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class TimeProfile(BaseModel):
+    profile_id: str = Field(pattern=r"^[a-z0-9_-]{1,64}$")
+    label: str = Field(min_length=1, max_length=80)
+    sample_count: int = Field(default=0, ge=0)
+    dominant_state: str | None = None
+    confidence: float = Field(default=0.0, ge=0.0, le=1.0)
+
+
 class RelatedAutomation(BaseModel):
    entity_id: str = Field(pattern=r"^automation\.[a-z0-9_]+$")
    config_id: str = Field(min_length=1, max_length=120)
@@ -150,6 +255,21 @@ class BehaviorState(BaseModel):
    related_automations: list[RelatedAutomation] = Field(default_factory=list)
    paused_automation_entity_ids: list[str] = Field(default_factory=list)
    reason: str = "Historische Aktorhandlungen werden analysiert."
+    safety: SafetyProfile = Field(default_factory=SafetyProfile)
+    decision_factors: list[DecisionFactor] = Field(default_factory=list)
+    knowledge: list[str] = Field(default_factory=list)
+    assumptions: list[str] = Field(default_factory=list)
+    uncertainties: list[str] = Field(default_factory=list)
+    safety_blockers: list[str] = Field(default_factory=list)
+    sample_trend: list[int] = Field(default_factory=list)
+    confidence_trend: list[float] = Field(default_factory=list)
+    correct_feedback_count: int = Field(default=0, ge=0)
+    incorrect_feedback_count: int = Field(default=0, ge=0)
+    model_snapshots: list[ModelSnapshot] = Field(default_factory=list)
+    active_model_version: str | None = None
+    adaptive_weight_updates: list[AdaptiveWeightUpdate] = Field(default_factory=list)
+    automation_conflicts: list[AutomationConflict] = Field(default_factory=list)
+    time_profiles: list[TimeProfile] = Field(default_factory=list)


 class ActuatorRecord(BaseModel):
@@ -176,5 +296,22 @@ class ReconciliationState(BaseModel):
    last_summary: str = "Noch keine Reconciliation ausgeführt."


+class JobQueueItem(BaseModel):
+    job_id: str = Field(min_length=1, max_length=120)
+    kind: str = Field(min_length=1, max_length=40)
+    target: str | None = Field(default=None, max_length=160)
+    trigger: str = Field(default="manual", max_length=40)
+    status: JobStatus = JobStatus.PENDING
+    started_at: datetime | None = None
+    completed_at: datetime | None = None
+    duration_ms: int | None = Field(default=None, ge=0)
+    error: str | None = Field(default=None, max_length=500)
+    summary: str = Field(default="", max_length=500)
+
+
+class JobQueueState(BaseModel):
+    jobs: list[JobQueueItem] = Field(default_factory=list)
+
+
 def model_id_for_actuator(actuator_entity_id: str) -> str:
    return f"actuator.{actuator_entity_id}"
--- a/app/actuators/store.py
+++ b/app/actuators/store.py
@@ -8,6 +8,9 @@ from threading import RLock

 from app.actuators.models import (
    ActuatorRecord,
+    JobQueueItem,
+    JobQueueState,
+    JobStatus,
    LifecycleStatus,
    ModelLifecycleState,
    ReconciliationState,
@@ -22,6 +25,7 @@ class ActuatorStore:
        self._actuators_root.mkdir(parents=True, exist_ok=True)
        self._lock = RLock()
        self._reconciliation_state_path = self._root / "reconciliation_state.json"
+        self._job_queue_path = self._root / "job_queue.json"

    def list(self) -> list[ActuatorRecord]:
        with self._lock:
@@ -85,6 +89,75 @@ class ActuatorStore:
            self._persist_reconciliation_state(state)
            return state

+    def load_job_queue(self) -> JobQueueState:
+        with self._lock:
+            if not self._job_queue_path.exists():
+                return JobQueueState()
+            try:
+                return JobQueueState.model_validate_json(
+                    self._job_queue_path.read_text(encoding="utf-8")
+                )
+            except ValueError as exc:
+                raise ValueError("Ungültiger Job-Queue-Status.") from exc
+
+    def start_job(
+        self,
+        *,
+        kind: str,
+        trigger: str,
+        target: str | None = None,
+        summary: str = "",
+    ) -> JobQueueItem:
+        now = datetime.now(timezone.utc)
+        job = JobQueueItem(
+            job_id=f"{now.strftime('%Y%m%d%H%M%S%f')}-{kind}-{target or 'all'}",
+            kind=kind,
+            target=target,
+            trigger=trigger,
+            status=JobStatus.RUNNING,
+            started_at=now,
+            summary=summary,
+        )
+        with self._lock:
+            queue = self.load_job_queue()
+            queue.jobs = [*queue.jobs, job][-50:]
+            self._persist_job_queue(queue)
+        return job
+
+    def finish_job(
+        self,
+        job_id: str,
+        *,
+        status: JobStatus,
+        summary: str = "",
+        error: str | None = None,
+    ) -> JobQueueItem | None:
+        now = datetime.now(timezone.utc)
+        with self._lock:
+            queue = self.load_job_queue()
+            updated_job: JobQueueItem | None = None
+            jobs: list[JobQueueItem] = []
+            for job in queue.jobs:
+                if job.job_id != job_id:
+                    jobs.append(job)
+                    continue
+                duration_ms = None
+                if job.started_at is not None:
+                    duration_ms = max(0, int((now - job.started_at).total_seconds() * 1000))
+                updated_job = job.model_copy(
+                    update={
+                        "status": status,
+                        "completed_at": now,
+                        "duration_ms": duration_ms,
+                        "summary": summary or job.summary,
+                        "error": error,
+                    }
+                )
+                jobs.append(updated_job)
+            queue.jobs = jobs[-50:]
+            self._persist_job_queue(queue)
+            return updated_job
+
    def _target(self, actuator_entity_id: str) -> Path:
        if "." not in actuator_entity_id:
            raise ValueError("Ungültige actuator_entity_id.")
@@ -108,6 +181,14 @@ class ActuatorStore:
        )
        os.replace(temporary, self._reconciliation_state_path)

+    def _persist_job_queue(self, state: JobQueueState) -> None:
+        temporary = self._job_queue_path.with_suffix(".json.tmp")
+        temporary.write_text(
+            json.dumps(state.model_dump(mode="json"), ensure_ascii=True, sort_keys=True) + "\n",
+            encoding="utf-8",
+        )
+        os.replace(temporary, self._job_queue_path)
+
    @staticmethod
    def _load(path: Path) -> ActuatorRecord:
        try:
--- a/app/api/v1/actuators.py
+++ b/app/api/v1/actuators.py
@@ -10,6 +10,7 @@ from pydantic import BaseModel, Field

 from app.actuators.lifecycle import ActuatorReconciliationService
 from app.actuators.models import ActuatorRecord, ReconciliationState, SensorWeightGroup
+from app.actuators.models import JobQueueItem, JobQueueState, JobStatus, SafetyProfile
 from app.actuators.store import ActuatorStore
 from app.behavior.engine import BehaviorEngine
 from app.config import Settings
@@ -55,6 +56,14 @@ class FeedbackRequest(BaseModel):
    expected_state: str | None = Field(default=None, max_length=100)


+class SafetyProfileRequest(BaseModel):
+    safety: SafetyProfile
+
+
+class ModelRollbackRequest(BaseModel):
+    version_id: str = Field(min_length=1, max_length=120)
+
+
 class ActuatorSuggestion(BaseModel):
    entity_id: str
    domain: str
@@ -112,6 +121,7 @@ class DashboardOverview(BaseModel):
    cache: EntityCacheStatus
    actuators: list[ActuatorSummary]
    discovery_groups: list[DashboardDiscoveryGroup]
+    jobs: JobQueueState = Field(default_factory=JobQueueState)


@router.get("/discovery", response_model=list[HaEntitySummary])
@@ -124,8 +134,19 @@ def discover_actuators(
    if cached_entities:
        entities = {entity.entity_id: entity for entity in cached_entities}
    else:
-        fresh_entities = list(ha_reader.read_entities())
-        _save_cached_entities(request, fresh_entities)
+        job = _start_job(
+            request,
+            kind="discovery",
+            trigger="manual" if refresh else "cache-miss",
+            summary="Home-Assistant-Entities werden gelesen und klassifiziert.",
+        )
+        try:
+            fresh_entities = list(ha_reader.read_entities())
+            _save_cached_entities(request, fresh_entities)
+        except Exception as exc:
+            _finish_job(job, request, status=JobStatus.FAILED, summary="Discovery fehlgeschlagen.", error=str(exc))
+            raise
+        _finish_job(job, request, status=JobStatus.COMPLETED, summary=f"{len(fresh_entities)} Entities klassifiziert.")
        entities = {entity.entity_id: entity for entity in fresh_entities}
    discovered = discover_entities(list(entities.values()))
    actuator_ids = _deduplicate_actuator_ids(
@@ -278,6 +299,12 @@ def dashboard_overview(request: Request) -> DashboardOverview:
    reconciliation = _reconciliation_state_or_default(request)
    ws_status = getattr(request.app.state, "ws_status", None)
    actuators = list_configured_summary(request)
+    store = getattr(request.app.state, "actuator_store", None)
+    jobs = (
+        store.load_job_queue()
+        if isinstance(store, ActuatorStore)
+        else JobQueueState()
+    )
    return DashboardOverview(
        system=DashboardSystemStatus(
            websocket_status=getattr(ws_status, "status", "unavailable"),
@@ -298,6 +325,7 @@ def dashboard_overview(request: Request) -> DashboardOverview:
        ),
        actuators=actuators,
        discovery_groups=cached_groups,
+        jobs=jobs,
    )


@@ -372,6 +400,32 @@ def record_feedback(
        raise HTTPException(status_code=404, detail=str(exc)) from exc


+@router.post("/{actuator_entity_id}/safety", response_model=ActuatorRecord)
+def set_safety_profile(
+    actuator_entity_id: str,
+    payload: SafetyProfileRequest,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).set_safety_profile(actuator_entity_id, profile=payload.safety)
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+
+@router.post("/{actuator_entity_id}/model/rollback", response_model=ActuatorRecord)
+def rollback_model(
+    actuator_entity_id: str,
+    payload: ModelRollbackRequest,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).rollback_model(actuator_entity_id, version_id=payload.version_id)
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    except ValueError as exc:
+        raise HTTPException(status_code=422, detail=str(exc)) from exc
+
+
@router.post("/{actuator_entity_id}/activation", response_model=ActuatorRecord)
 def set_activation(
    actuator_entity_id: str,
@@ -441,11 +495,27 @@ def refresh_related_automations(
    actuator_entity_id: str,
    request: Request,
 ) -> ActuatorRecord:
+    job = _start_job(
+        request,
+        kind="automation_refresh",
+        trigger="manual",
+        target=actuator_entity_id,
+        summary="Passende HA-Automationen werden gesucht.",
+    )
    try:
-        return _behavior(request).refresh_related_automations(actuator_entity_id)
+        record = _behavior(request).refresh_related_automations(actuator_entity_id)
+        _finish_job(
+            job,
+            request,
+            status=JobStatus.COMPLETED,
+            summary=f"{len(record.behavior.related_automations)} Automationen gefunden.",
+        )
+        return record
    except KeyError as exc:
+        _finish_job(job, request, status=JobStatus.FAILED, summary="Automation-Refresh fehlgeschlagen.", error=str(exc))
        raise HTTPException(status_code=404, detail=str(exc)) from exc
    except (ValueError, HaClientError) as exc:
+        _finish_job(job, request, status=JobStatus.FAILED, summary="Automation-Refresh fehlgeschlagen.", error=str(exc))
        raise HTTPException(status_code=409, detail=str(exc)) from exc


@@ -486,12 +556,85 @@ def run_reconciliation(
    request: Request,
    trigger: str = Query(default="manual", pattern=r"^[a-z0-9_-]{1,32}$"),
 ) -> ReconciliationState:
-    state = _service(request).reconcile_all(trigger=trigger)
-    _behavior(request).train_all()
-    _behavior(request).evaluate_all()
+    reconciliation_job = _start_job(
+        request,
+        kind="reconciliation",
+        trigger=trigger,
+        summary="Kontext, Zuordnung und Modelle werden abgeglichen.",
+    )
+    training_job: JobQueueItem | None = None
+    evaluation_job: JobQueueItem | None = None
+    try:
+        state = _service(request).reconcile_all(trigger=trigger)
+        _finish_job(reconciliation_job, request, status=JobStatus.COMPLETED, summary=state.last_summary)
+        reconciliation_job = None
+        training_job = _start_job(
+            request,
+            kind="training",
+            trigger=trigger,
+            summary="Gelernte Aktorhandlungen werden aktualisiert.",
+        )
+        _behavior(request).train_all()
+        _finish_job(training_job, request, status=JobStatus.COMPLETED, summary="Training abgeschlossen.")
+        training_job = None
+        evaluation_job = _start_job(
+            request,
+            kind="evaluation",
+            trigger=trigger,
+            summary="Aktuelle Vorhersagen werden neu berechnet.",
+        )
+        _behavior(request).evaluate_all()
+        _finish_job(evaluation_job, request, status=JobStatus.COMPLETED, summary="Evaluation abgeschlossen.")
+        evaluation_job = None
+    except Exception as exc:
+        for job in [reconciliation_job, training_job, evaluation_job]:
+            if isinstance(job, JobQueueItem) and job.status is JobStatus.RUNNING:
+                _finish_job(job, request, status=JobStatus.FAILED, summary="Job fehlgeschlagen.", error=str(exc))
+        raise
    return state


+@router.get("/job-queue/state", response_model=JobQueueState)
+def get_job_queue(request: Request) -> JobQueueState:
+    store = getattr(request.app.state, "actuator_store", None)
+    if not isinstance(store, ActuatorStore):
+        raise HTTPException(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            detail="Actuator Store nicht initialisiert.",
+        )
+    return store.load_job_queue()
+
+
+def _start_job(
+    request: Request,
+    *,
+    kind: str,
+    trigger: str,
+    target: str | None = None,
+    summary: str = "",
+) -> JobQueueItem | None:
+    store = getattr(request.app.state, "actuator_store", None)
+    if not isinstance(store, ActuatorStore):
+        return None
+    return store.start_job(kind=kind, trigger=trigger, target=target, summary=summary)
+
+
+def _finish_job(
+    job: JobQueueItem | None,
+    request: Request,
+    *,
+    status: JobStatus,
+    summary: str,
+    error: str | None = None,
+) -> None:
+    if job is None:
+        return
+    store = getattr(request.app.state, "actuator_store", None)
+    if not isinstance(store, ActuatorStore):
+        return
+    store.finish_job(job.job_id, status=status, summary=summary, error=error)
+
+
 def _service(request: Request) -> ActuatorReconciliationService:
    service = getattr(request.app.state, "actuator_service", None)
    if not isinstance(service, ActuatorReconciliationService):
--- a/app/behavior/engine.py
+++ b/app/behavior/engine.py
@@ -7,13 +7,21 @@ from zoneinfo import ZoneInfo

 from app.actuators.models import (
    ActuatorRecord,
+    AdaptiveWeightUpdate,
+    AutomationConflict,
    BehaviorMode,
    BehaviorPattern,
    BehaviorPrediction,
    BehaviorState,
    BehaviorStatus,
+    DecisionFactor,
    ExecutionEvent,
+    ManualOverride,
+    ModelSnapshot,
    RelatedAutomation,
+    SafetyProfile,
+    SafetyStage,
+    TimeProfile,
 )
 from app.actuators.store import ActuatorStore
 from app.config import Settings
@@ -165,6 +173,7 @@ class BehaviorEngine:
                "eindeutig zugeordnete Handlungen fehlen."
            )
        )
+        model_version_id = f"model-{now.strftime('%Y%m%d%H%M%S')}"
        behavior = record.behavior.model_copy(
            update={
                "status": status,
@@ -175,6 +184,22 @@ class BehaviorEngine:
                "patterns": patterns[-_MAX_PATTERNS:],
                "last_trained_at": now,
                "reason": reason,
+                "sample_trend": [*record.behavior.sample_trend, len(patterns)][-30:],
+                "knowledge": _knowledge_lines(record, len(patterns), trusted_actions),
+                "assumptions": _assumption_lines(record),
+                "uncertainties": _uncertainty_lines(record, len(patterns), trusted_actions),
+                "time_profiles": _time_profiles(patterns),
+                "model_snapshots": _next_model_snapshots(
+                    record.behavior.model_snapshots,
+                    model_version_id,
+                    patterns[-_MAX_PATTERNS:],
+                    len(patterns),
+                    trusted_actions,
+                    _average(record.behavior.confidence_trend),
+                    record.behavior.incorrect_feedback_count,
+                    reason,
+                ),
+                "active_model_version": model_version_id,
            }
        )
        return self._save_behavior(record, behavior)
@@ -268,16 +293,25 @@ class BehaviorEngine:
            timezone_name=self._settings.timezone,
        )
        if prediction is not None:
+            safety_allowed, safety_blockers = self._assess_safety(
+                record,
+                actuator.state,
+                prediction,
+                now,
+            )
            prediction = prediction.model_copy(
                update={
-                    "execution_reason": self._prediction_execution_reason(
-                        record,
-                        actuator.state,
-                        prediction,
-                        now,
+                    "execution_reason": (
+                        "Ausführung ist freigegeben."
+                        if safety_allowed
+                        else "Nicht ausgeführt: " + " ".join(safety_blockers)
                    )
                }
            )
+        else:
+            safety_allowed = False
+            safety_blockers = ["Keine fällige Vorhersage."]
+        decision_factors = _decision_factors_for(record, current_context, prediction)
        behavior = record.behavior.model_copy(
            update={
                "last_evaluated_at": now,
@@ -287,18 +321,21 @@ class BehaviorEngine:
                    if prediction is not None
                    else "Aktuell ist kein gelerntes Handlungsmuster fällig."
                ),
+                "decision_factors": decision_factors,
+                "knowledge": _knowledge_lines(record, record.behavior.sample_count, record.behavior.high_confidence_sample_count),
+                "assumptions": _assumption_lines(record),
+                "uncertainties": _uncertainty_lines(record, record.behavior.sample_count, record.behavior.high_confidence_sample_count),
+                "safety_blockers": safety_blockers if prediction is not None else [],
+                "confidence_trend": (
+                    [*record.behavior.confidence_trend, round(prediction.confidence, 4)][-30:]
+                    if prediction is not None
+                    else record.behavior.confidence_trend
+                ),
            }
        )
        if (
            prediction is not None
-            and behavior.mode is BehaviorMode.ACTIVE
-            and prediction.confidence >= self._settings.prediction_confidence
-            and actuator.state != prediction.target_state
-            and self._cooldown_elapsed(
-                behavior,
-                now,
-                prediction.target_state,
-            )
+            and safety_allowed
        ):
            domain = actuator_entity_id.split(".", 1)[0]
            service = service_for_state(domain, prediction.target_state)
@@ -403,6 +440,8 @@ class BehaviorEngine:
                )
            )
            reason = "Vorhersage wurde vom Nutzer als korrekt bestätigt."
+            correct_count = record.behavior.correct_feedback_count + 1
+            incorrect_count = record.behavior.incorrect_feedback_count
        else:
            target = prediction.target_state if prediction is not None else None
            if target:
@@ -431,6 +470,13 @@ class BehaviorEngine:
                    )
                )
            reason = "Vorhersage wurde vom Nutzer als falsch markiert."
+            correct_count = record.behavior.correct_feedback_count
+            incorrect_count = record.behavior.incorrect_feedback_count + 1
+        adaptive_updates, manual_override = _adapt_sensor_weights(
+            record,
+            current_context,
+            correct=correct,
+        )
        behavior = record.behavior.model_copy(
            update={
                "patterns": patterns[-_MAX_PATTERNS:],
@@ -441,6 +487,56 @@ class BehaviorEngine:
                ),
                "reason": reason,
                "last_trained_at": now,
+                "correct_feedback_count": correct_count,
+                "incorrect_feedback_count": incorrect_count,
+                "adaptive_weight_updates": [
+                    *record.behavior.adaptive_weight_updates,
+                    *adaptive_updates,
+                ][-50:],
+            }
+        )
+        record_for_save = (
+            record.model_copy(update={"manual_override": manual_override})
+            if manual_override is not None
+            else record
+        )
+        return self._save_behavior(record_for_save, behavior)
+
+    def rollback_model(
+        self,
+        actuator_entity_id: str,
+        *,
+        version_id: str,
+    ) -> ActuatorRecord:
+        record = self._store.get(actuator_entity_id)
+        snapshot = next(
+            (item for item in record.behavior.model_snapshots if item.version_id == version_id),
+            None,
+        )
+        if snapshot is None:
+            raise ValueError("Modell-Snapshot nicht gefunden.")
+        behavior = record.behavior.model_copy(
+            update={
+                "patterns": snapshot.patterns,
+                "sample_count": snapshot.sample_count,
+                "high_confidence_sample_count": snapshot.high_confidence_sample_count,
+                "active_model_version": snapshot.version_id,
+                "reason": f"Rollback auf Modell-Snapshot {snapshot.version_id}.",
+            }
+        )
+        return self._save_behavior(record, behavior)
+
+    def set_safety_profile(
+        self,
+        actuator_entity_id: str,
+        *,
+        profile: SafetyProfile,
+    ) -> ActuatorRecord:
+        record = self._store.get(actuator_entity_id)
+        behavior = record.behavior.model_copy(
+            update={
+                "safety": profile.model_copy(update={"updated_at": datetime.now(timezone.utc)}),
+                "reason": "Sicherheitsprofil wurde manuell aktualisiert.",
            }
        )
        return self._save_behavior(record, behavior)
@@ -459,7 +555,10 @@ class BehaviorEngine:
            )
        ]
        behavior = record.behavior.model_copy(
-            update={"related_automations": related}
+            update={
+                "related_automations": related,
+                "automation_conflicts": _automation_conflicts(record, related),
+            }
        )
        return self._save_behavior(record, behavior)

@@ -527,6 +626,9 @@ class BehaviorEngine:
                update={
                    "mode": mode,
                    "approved_at": approved_at,
+                    "safety": record.behavior.safety.model_copy(
+                        update={"stage": SafetyStage.ACTIVE, "updated_at": now}
+                    ),
                    "reason": (
                        "Autonomes Lernen und Schalten wurde ausdrücklich freigegeben."
                    ),
@@ -607,6 +709,9 @@ class BehaviorEngine:
            update={
                "mode": mode,
                "approved_at": approved_at,
+                "safety": record.behavior.safety.model_copy(
+                    update={"stage": SafetyStage.SHADOW, "updated_at": now}
+                ),
                "related_automations": [
                    automation.model_copy(update={"enabled": True})
                    if (
@@ -651,6 +756,51 @@ class BehaviorEngine:
            return "Nicht ausgeführt: Sicherheits-Cooldown ist noch aktiv."
        return "Ausführung ist freigegeben."

+    def _assess_safety(
+        self,
+        record: ActuatorRecord,
+        current_state: str | None,
+        prediction: BehaviorPrediction,
+        now: datetime,
+    ) -> tuple[bool, list[str]]:
+        profile = record.behavior.safety
+        blockers: list[str] = []
+        domain = record.actuator_entity_id.split(".", 1)[0]
+        if not record.enabled:
+            blockers.append("Aktor ist in SillyHome deaktiviert.")
+        if domain not in _SAFE_ACTIVE_DOMAINS:
+            blockers.append(f"Domain {domain} ist nicht für autonomes Schalten freigegeben.")
+        if profile.manual_block:
+            blockers.append("Manuelle Sicherheitssperre ist aktiv.")
+        stage = profile.stage
+        if (
+            record.behavior.mode is BehaviorMode.ACTIVE
+            and profile.updated_at is None
+            and stage is SafetyStage.SHADOW
+        ):
+            stage = SafetyStage.ACTIVE
+        if stage not in {SafetyStage.ACTIVE, SafetyStage.PARTIAL}:
+            blockers.append(f"Safety-Stufe {stage.value} erlaubt noch kein Schalten.")
+        if record.behavior.mode is not BehaviorMode.ACTIVE:
+            blockers.append("SillyHome ist im Shadow-Modus.")
+        if not record.behavior.activation_ready:
+            blockers.append(record.behavior.activation_reason)
+        threshold = _confidence_threshold_for(profile, prediction.target_state)
+        if prediction.confidence < threshold:
+            blockers.append(
+                f"Sicherheit {prediction.confidence:.0%} liegt unter der Schwelle {threshold:.0%}."
+            )
+        if current_state == prediction.target_state:
+            blockers.append("Zielzustand ist bereits erreicht.")
+        if not self._cooldown_elapsed(
+            record.behavior,
+            now,
+            prediction.target_state,
+            cooldown_seconds=profile.cooldown_seconds,
+        ):
+            blockers.append("Sicherheits-Cooldown ist noch aktiv.")
+        return not blockers, blockers
+
    def _build_patterns(
        self,
        *,
@@ -701,6 +851,8 @@ class BehaviorEngine:
        behavior: BehaviorState,
        now: datetime,
        target_state: str,
+        *,
+        cooldown_seconds: int | None = None,
    ) -> bool:
        if behavior.last_executed_at is None:
            return True
@@ -708,7 +860,9 @@ class BehaviorEngine:
        if last_event is not None and last_event.target_state != target_state:
            return True
        return (now - behavior.last_executed_at) >= timedelta(
-            seconds=self._settings.execution_cooldown_seconds
+            seconds=cooldown_seconds
+            if cooldown_seconds is not None
+            else self._settings.execution_cooldown_seconds
        )

    def _save_behavior(
@@ -791,6 +945,269 @@ def _event_changed_at(new_state: dict[str, object] | None) -> datetime | None:
    return parsed


+def _confidence_threshold_for(profile: SafetyProfile, target_state: str) -> float:
+    if target_state == "on" and profile.min_confidence_on is not None:
+        return profile.min_confidence_on
+    if target_state in {"off", "closed"} and profile.min_confidence_off is not None:
+        return profile.min_confidence_off
+    return profile.min_confidence
+
+
+def _decision_factors_for(
+    record: ActuatorRecord,
+    current_context: dict[str, str | None],
+    prediction: BehaviorPrediction | None,
+) -> list[DecisionFactor]:
+    factors: list[DecisionFactor] = []
+    candidates = {
+        candidate.entity_id: candidate
+        for candidate in [*record.numeric_candidates, *record.context_candidates]
+    }
+    for entity_id, state in current_context.items():
+        candidate = candidates.get(entity_id)
+        weight = candidate.effective_weight if candidate is not None else 1.0
+        relevance = candidate.confidence if candidate is not None else 0.5
+        contribution = round(min(1.0, weight * relevance), 4)
+        factors.append(
+            DecisionFactor(
+                entity_id=entity_id,
+                label=(
+                    candidate.friendly_name
+                    if candidate is not None and candidate.friendly_name
+                    else entity_id
+                ),
+                factor_type="context",
+                state=state,
+                weight=round(weight, 4),
+                contribution=contribution,
+                evidence=(
+                    candidate.evidence[:4]
+                    if candidate is not None
+                    else ["Aktuell ausgewähltes Kontextsignal."]
+                ),
+            )
+        )
+    if prediction is not None:
+        factors.append(
+            DecisionFactor(
+                label=f"Vorhersage {prediction.target_state}",
+                factor_type="prediction",
+                state=prediction.target_state,
+                weight=1.0,
+                contribution=prediction.confidence,
+                evidence=[prediction.reason],
+            )
+        )
+    return sorted(factors, key=lambda item: (-item.contribution, item.label))[:12]
+
+
+def _knowledge_lines(
+    record: ActuatorRecord,
+    sample_count: int,
+    trusted_actions: int,
+) -> list[str]:
+    lines = [
+        f"{sample_count} historische Aktorhandlungen sind ausgewertet.",
+        f"{trusted_actions} Handlungen stammen eindeutig von Nutzer oder HA-Automationen.",
+    ]
+    if record.assignment.selected_numeric_entity_id:
+        lines.append(f"Hauptsensor: {record.assignment.selected_numeric_entity_id}.")
+    if record.assignment.selected_context_entity_ids:
+        lines.append(
+            f"{len(record.assignment.selected_context_entity_ids)} Kontextsignale sind verbunden."
+        )
+    return lines
+
+
+def _assumption_lines(record: ActuatorRecord) -> list[str]:
+    lines = [
+        "Ähnliche Zeitfenster und ähnliche Kontextzustände deuten auf ähnliche Nutzerabsicht hin."
+    ]
+    if record.manual_override is not None:
+        lines.append("Manuelle Sensor-/Kontextkorrekturen werden höher gewichtet.")
+    if record.behavior.related_automations:
+        lines.append("Passende HA-Automationen gelten als starker Hinweis auf vorhandene Logik.")
+    return lines
+
+
+def _uncertainty_lines(
+    record: ActuatorRecord,
+    sample_count: int,
+    trusted_actions: int,
+) -> list[str]:
+    lines: list[str] = []
+    if sample_count < trusted_actions + 3:
+        lines.append("Noch wenig Varianz in den gelernten Handlungen.")
+    if trusted_actions < sample_count:
+        lines.append("Ein Teil der Handlungen ist nicht eindeutig Nutzer oder Automation zugeordnet.")
+    if record.assignment.review_required:
+        lines.append("Die automatische Kontextzuordnung verlangt noch Prüfung.")
+    if record.behavior.incorrect_feedback_count:
+        lines.append(
+            f"{record.behavior.incorrect_feedback_count} negative Feedbacks senken Vertrauen."
+        )
+    return lines or ["Keine kritische Unsicherheit aus den lokalen Daten erkannt."]
+
+
+def _next_model_snapshots(
+    existing: list[ModelSnapshot],
+    version_id: str,
+    patterns: list[BehaviorPattern],
+    sample_count: int,
+    trusted_actions: int,
+    average_confidence: float,
+    incorrect_feedback_count: int,
+    reason: str,
+) -> list[ModelSnapshot]:
+    snapshot = ModelSnapshot(
+        version_id=version_id,
+        sample_count=sample_count,
+        high_confidence_sample_count=trusted_actions,
+        average_confidence=round(average_confidence, 4),
+        incorrect_feedback_count=incorrect_feedback_count,
+        patterns=patterns,
+        reason=reason,
+    )
+    return [*existing, snapshot][-10:]
+
+
+def _average(values: list[float]) -> float:
+    return sum(values) / len(values) if values else 0.0
+
+
+def _time_profiles(patterns: list[BehaviorPattern]) -> list[TimeProfile]:
+    buckets = {
+        "night": ("Nacht", range(0, 360)),
+        "morning": ("Morgen", range(360, 720)),
+        "day": ("Tag", range(720, 1080)),
+        "evening": ("Abend", range(1080, 1440)),
+    }
+    profiles: list[TimeProfile] = []
+    for profile_id, (label, minutes) in buckets.items():
+        selected = [pattern for pattern in patterns if pattern.minute_of_day in minutes]
+        if not selected:
+            profiles.append(TimeProfile(profile_id=profile_id, label=label))
+            continue
+        by_state: dict[str, int] = {}
+        for pattern in selected:
+            by_state[pattern.target_state] = by_state.get(pattern.target_state, 0) + 1
+        dominant_state, count = max(by_state.items(), key=lambda item: (item[1], item[0]))
+        profiles.append(
+            TimeProfile(
+                profile_id=profile_id,
+                label=label,
+                sample_count=len(selected),
+                dominant_state=dominant_state,
+                confidence=round(count / len(selected), 4),
+            )
+        )
+    weekend = [pattern for pattern in patterns if pattern.weekday >= 5]
+    profiles.append(
+        TimeProfile(
+            profile_id="weekend",
+            label="Wochenende",
+            sample_count=len(weekend),
+            dominant_state=(
+                max(
+                    {pattern.target_state: 0 for pattern in weekend},
+                    key=lambda state: sum(pattern.target_state == state for pattern in weekend),
+                )
+                if weekend
+                else None
+            ),
+            confidence=round(len(weekend) / len(patterns), 4) if patterns else 0.0,
+        )
+    )
+    return profiles
+
+
+def _adapt_sensor_weights(
+    record: ActuatorRecord,
+    current_context: dict[str, str | None],
+    *,
+    correct: bool,
+) -> tuple[list[AdaptiveWeightUpdate], ManualOverride | None]:
+    if not current_context:
+        return [], record.manual_override
+    candidates = {
+        candidate.entity_id: candidate
+        for candidate in [*record.numeric_candidates, *record.context_candidates]
+    }
+    previous = record.manual_override
+    weights = dict(previous.sensor_weights if previous is not None else {})
+    updates: list[AdaptiveWeightUpdate] = []
+    delta = 0.03 if correct else -0.08
+    for entity_id in current_context:
+        candidate = candidates.get(entity_id)
+        base = weights.get(
+            entity_id,
+            candidate.effective_weight if candidate is not None else 1.0,
+        )
+        new_weight = round(min(1.0, max(0.1, base + delta)), 4)
+        if new_weight == base:
+            continue
+        weights[entity_id] = new_weight
+        updates.append(
+            AdaptiveWeightUpdate(
+                entity_id=entity_id,
+                previous_weight=round(base, 4),
+                new_weight=new_weight,
+                reason=(
+                    "Feedback korrekt: Kontextsignal leicht höher gewichtet."
+                    if correct
+                    else "Feedback falsch: Kontextsignal vorsichtig abgewertet."
+                ),
+            )
+        )
+    if not updates:
+        return [], previous
+    return updates, ManualOverride(
+        numeric_entity_id=(
+            previous.numeric_entity_id
+            if previous is not None
+            else record.assignment.selected_numeric_entity_id
+        ),
+        context_entity_ids=(
+            previous.context_entity_ids
+            if previous is not None
+            else record.assignment.selected_context_entity_ids
+        ),
+        sensor_weights=weights,
+        sensor_weight_groups=previous.sensor_weight_groups if previous is not None else [],
+        note="Sensor-Gewichtungen automatisch aus Feedback angepasst.",
+    )
+
+
+def _automation_conflicts(
+    record: ActuatorRecord,
+    related: list[RelatedAutomation],
+) -> list[AutomationConflict]:
+    conflicts: list[AutomationConflict] = []
+    for automation in related:
+        if record.behavior.mode is BehaviorMode.ACTIVE and automation.enabled:
+            conflicts.append(
+                AutomationConflict(
+                    automation_entity_id=automation.entity_id,
+                    severity="warning",
+                    status="open",
+                    reason=(
+                        "SillyHome ist aktiv, aber diese passende HA-Automation "
+                        "ist ebenfalls aktiv. Das kann zu konkurrierenden Schaltungen führen."
+                    ),
+                )
+            )
+        elif automation.entity_id in record.behavior.paused_automation_entity_ids:
+            conflicts.append(
+                AutomationConflict(
+                    automation_entity_id=automation.entity_id,
+                    severity="info",
+                    status="controlled",
+                    reason="Automation ist durch SillyHome pausiert.",
+                )
+            )
+    return conflicts
+
+
 def predict_behavior(
    patterns: list[BehaviorPattern],
    *,
--- a/app/main.py
+++ b/app/main.py
@@ -105,7 +105,7 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
 app = FastAPI(
    title="SillyHome Next API",
    description="Lokales Smart-Home-Intelligenzsystem für Home Assistant.",
-    version="1.0.4",
+    version="1.2.0",
    lifespan=lifespan,
 )
 app.state.settings = load_settings()
--- a/app/static/index.html
+++ b/app/static/index.html
@@ -25,6 +25,7 @@
      --bad:#8fb8ff;
    }
    * { box-sizing:border-box; }
+    html, body { max-width:100%; overflow-x:hidden; }
    body { margin:0; font-size:15px; background:var(--panel-quiet); }
    h1,h2,h3 { margin:0 0 10px; letter-spacing:0; }
    p { margin:6px 0; }
@@ -71,31 +72,38 @@
    .bad { color: var(--bad); }
    label { display:block; margin:9px 0 4px; color:#c3d2df; font-weight:700; }
    select,input,button { width:100%; border-radius:8px; border:1px solid #3b4b5b; padding:10px; background:#111821; color:#fff; font:inherit; min-width:0; }
+    input[type="checkbox"] { width:auto; min-width:0; vertical-align:middle; margin-right:8px; }
    select[multiple] { min-height:150px; }
    button { min-height:42px; margin-top:10px; background:var(--accent); color:#211204; border:0; font-weight:850; cursor:pointer; }
    button.secondary { background:var(--complement-soft); color:#dff6ff; border:1px solid #22607c; }
    button.danger { background:#2a3441; color:#f2f6fb; border:1px solid #536273; }
    button.compact { width:auto; min-width:112px; margin-right:8px; padding:8px 10px; min-height:36px; }
-    table { width: 100%; border-collapse: collapse; font-size: .92rem; }
-    td,th { padding: 8px; border-bottom: 1px solid #2d3a47; text-align: left; vertical-align: top; }
+    table { width: 100%; border-collapse: collapse; table-layout:fixed; font-size: .92rem; }
+    td,th { padding: 8px; border-bottom: 1px solid #2d3a47; text-align: left; vertical-align: top; overflow-wrap:anywhere; word-break:break-word; }
    ul { margin: 8px 0; padding-left: 18px; }
    .notice { border-left:4px solid var(--complement); padding-left:10px; }
    .grid-two { display:grid; grid-template-columns:repeat(auto-fit,minmax(240px,1fr)); gap:12px; }
    .chips { display:flex; flex-wrap:wrap; gap:6px; margin-top:8px; }
-    .chip { padding:4px 8px; border-radius:8px; background:#222b36; border:1px solid var(--border); font-size:.85rem; }
+    .chip { padding:4px 8px; border-radius:8px; background:#222b36; border:1px solid var(--border); font-size:.85rem; max-width:100%; overflow-wrap:anywhere; word-break:break-word; }
    .muted { color:var(--text-soft); }
    .card-list { display:grid; grid-template-columns:repeat(auto-fit,minmax(250px,1fr)); gap:10px; }
    .actuator-card { background:#121922; border:1px solid var(--border); border-radius:8px; padding:10px; min-width:0; }
    .actuator-card.selected { border-color:var(--complement); box-shadow:0 0 0 1px rgba(28,199,255,.35); }
-    .card-title { display:flex; justify-content:space-between; gap:10px; align-items:flex-start; margin-bottom:8px; }
-    .entity-id { overflow-wrap:anywhere; font-weight:800; }
+    .card-title { display:flex; flex-wrap:wrap; justify-content:space-between; gap:10px; align-items:flex-start; margin-bottom:8px; min-width:0; }
+    .card-title > div { min-width:0; flex:1 1 160px; overflow-wrap:anywhere; word-break:break-word; }
+    .card-title .chip { flex:0 1 auto; white-space:normal; text-align:center; }
+    .actuator-card strong { overflow-wrap:anywhere; word-break:break-word; }
+    .entity-id { overflow-wrap:anywhere; word-break:break-word; font-weight:800; }
    .metric-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(120px,1fr)); gap:6px; margin:8px 0; }
-    .metric { background:var(--panel-soft); border:1px solid var(--border); border-radius:8px; padding:8px; min-width:0; }
+    .metric { background:var(--panel-soft); border:1px solid var(--border); border-radius:8px; padding:8px; min-width:0; overflow-wrap:anywhere; word-break:break-word; }
    .metric strong { display:block; margin-bottom:4px; color:#cfe0ec; font-size:.84rem; }
+    .decision-list { display:grid; gap:8px; margin:10px 0; }
+    .decision-row { background:#121922; border:1px solid var(--border); border-radius:8px; padding:9px; min-width:0; overflow-wrap:anywhere; }
+    .decision-row header { padding:0; border:0; background:transparent; display:flex; justify-content:space-between; gap:10px; flex-wrap:wrap; }
    .actions { display:flex; flex-wrap:wrap; gap:8px; margin-top:10px; }
    .actions button { flex:1 1 180px; margin-top:0; }
    .detail-header { display:flex; justify-content:space-between; gap:12px; align-items:flex-start; flex-wrap:wrap; }
-    .manual-context { margin-top:12px; background:#121922; border:1px solid var(--border); border-radius:8px; padding:10px; }
+    .manual-context { margin-top:12px; background:#121922; border:1px solid var(--border); border-radius:8px; padding:10px; min-width:0; overflow-wrap:anywhere; word-break:break-word; }
    .inline-controls { display:grid; grid-template-columns:repeat(auto-fit,minmax(160px,1fr)); gap:8px; margin:8px 0; }
    .manual-entry { min-height:80px; resize:vertical; }
    textarea { width:100%; border-radius:8px; border:1px solid #3b4b5b; padding:12px; background:#101820; color:#fff; font:inherit; }
@@ -228,6 +236,7 @@
    <div id="status">Prüfung läuft ...</div>
    <div class="chips" id="status-chips"></div>
    <div id="dashboard-stats" class="metric-grid"></div>
+    <div id="job-queue" class="decision-list"></div>
  </section>

  <section class="guide-panel" id="guide">
@@ -482,6 +491,7 @@ function renderDashboardStatus(dashboard) {
  const status = document.getElementById("status");
  const chips = document.getElementById("status-chips");
  const stats = document.getElementById("dashboard-stats");
+  const jobsBox = document.getElementById("job-queue");
  const system = dashboard.system || {};
  const cache = dashboard.cache || {};
  const actuators = dashboard.actuators || [];
@@ -494,6 +504,8 @@ function renderDashboardStatus(dashboard) {
  ).length;
  const trainedCount = actuators.filter(record => record.behavior_status === "trained").length;
  const sampleTotal = actuators.reduce((sum, record) => sum + Number(record.sample_count || 0), 0);
+  const jobs = dashboard.jobs?.jobs || [];
+  const runningJobs = jobs.filter(job => job.status === "running").length;
  const cacheLabel = cache.available
    ? `Cache aktuell mit ${cache.entity_count} Entities`
    : "Cache wird nach Discovery aufgebaut";
@@ -509,6 +521,7 @@ function renderDashboardStatus(dashboard) {
    `<span class="chip">Aktoren: ${escapeHtml(system.configured_actuators ?? 0)}</span>`,
    `<span class="chip">Lernbereit: ${escapeHtml(system.trained_models ?? 0)}</span>`,
    `<span class="chip">Prüfen: ${escapeHtml(system.review_required ?? 0)}</span>`,
+    `<span class="chip">Jobs aktiv: ${escapeHtml(runningJobs)}</span>`,
  ].join("");
  stats.innerHTML = [
    `<div class="metric"><strong>Geladene Startdaten</strong>${escapeHtml(actuators.length)} Geräte</div>`,
@@ -519,6 +532,21 @@ function renderDashboardStatus(dashboard) {
    `<div class="metric"><strong>Discovery-Gruppen</strong>${escapeHtml(discoveryGroups.length)} Kategorien</div>`,
    `<div class="metric"><strong>Cache-Zeitpunkt</strong>${escapeHtml(cache.updated_at || "noch offen")}</div>`,
  ].join("");
+  jobsBox.innerHTML = jobs.length ? `
+    <h3>Job-Queue</h3>
+    ${jobs.slice(-6).reverse().map(job => `
+      <div class="decision-row">
+        <header>
+          <strong>${escapeHtml(job.kind)}${job.target ? `: ${escapeHtml(job.target)}` : ""}</strong>
+          <span class="chip">${escapeHtml(job.status)}</span>
+        </header>
+        <p class="muted">${escapeHtml(job.summary || "Keine Zusammenfassung")}</p>
+        <p class="muted">Start: ${escapeHtml(job.started_at || "offen")} · Dauer: ${escapeHtml(job.duration_ms == null ? "läuft/offen" : `${job.duration_ms} ms`)}</p>
+        ${job.error ? `<p class="bad">${escapeHtml(job.error)}</p>` : ""}
+        ${job.status === "failed" ? "<p class='warn'>Retry: Aktion im Dashboard erneut starten; der nächste Lauf schreibt einen neuen Queue-Eintrag.</p>" : ""}
+      </div>
+    `).join("")}
+  ` : "";
 }

 async function loadSummaryData() {
@@ -869,6 +897,113 @@ async function showActuator(actuatorId, evaluationMessage = "") {
        `).join("")}</ul>`
      : "<p class='muted'>Noch keine Kontext-Entity ausgewählt.</p>";
    const prediction = record.behavior.prediction;
+    const safety = record.behavior.safety || {};
+    const blockers = record.behavior.safety_blockers || [];
+    const decisionFactors = record.behavior.decision_factors || [];
+    const knowledge = record.behavior.knowledge || [];
+    const assumptions = record.behavior.assumptions || [];
+    const uncertainties = record.behavior.uncertainties || [];
+    const snapshots = record.behavior.model_snapshots || [];
+    const activeModelVersion = record.behavior.active_model_version || "";
+    const adaptiveUpdates = record.behavior.adaptive_weight_updates || [];
+    const automationConflicts = record.behavior.automation_conflicts || [];
+    const timeProfiles = record.behavior.time_profiles || [];
+    const safetyControls = `
+      <details class="manual-context" open>
+        <summary>Sicherheit und manuelles Gegensteuern</summary>
+        <div class="inline-controls">
+          <div>
+            <label for="safety-stage">Freigabestufe</label>
+            <select id="safety-stage">
+              ${["observe", "suggest", "shadow", "partial", "active"].map(stage => `
+                <option value="${stage}" ${safety.stage === stage ? "selected" : ""}>${stage}</option>
+              `).join("")}
+            </select>
+          </div>
+          <div>
+            <label for="safety-confidence">Mindest-Sicherheit in %</label>
+            <input id="safety-confidence" type="number" min="0" max="100" step="1" value="${Math.round((safety.min_confidence ?? 0.82) * 100)}">
+          </div>
+          <div>
+            <label for="safety-cooldown">Cooldown Sekunden</label>
+            <input id="safety-cooldown" type="number" min="0" step="10" value="${safety.cooldown_seconds ?? ""}" placeholder="Standard">
+          </div>
+        </div>
+        <label>
+          <input id="safety-manual-block" type="checkbox" ${safety.manual_block ? "checked" : ""}>
+          Manuelle Sicherheitssperre aktiv
+        </label>
+        ${blockers.length ? `<p class="warn">Aktuelle Blocker: ${blockers.map(escapeHtml).join(" ")}</p>` : "<p class='ok'>Keine lokalen Sicherheitsblocker für die aktuelle Vorhersage.</p>"}
+        <button class="secondary" onclick="saveSafetyProfile('${escapeHtml(record.actuator_entity_id)}')">Sicherheitsprofil speichern</button>
+      </details>
+    `;
+    const decisionArchive = `
+      <details class="manual-context" open>
+        <summary>Entscheidungsakte</summary>
+        <div class="grid-two">
+          <div>
+            <h3>Wissen</h3>
+            <ul>${knowledge.map(item => `<li>${escapeHtml(item)}</li>`).join("") || "<li>Keine gesicherten Punkte gespeichert.</li>"}</ul>
+          </div>
+          <div>
+            <h3>Annahmen</h3>
+            <ul>${assumptions.map(item => `<li>${escapeHtml(item)}</li>`).join("") || "<li>Keine Annahmen gespeichert.</li>"}</ul>
+          </div>
+        </div>
+        <h3>Unsicherheit</h3>
+        <ul>${uncertainties.map(item => `<li>${escapeHtml(item)}</li>`).join("") || "<li>Keine Unsicherheit gespeichert.</li>"}</ul>
+        <h3>Beitragsfaktoren</h3>
+        <div class="decision-list">
+          ${decisionFactors.length ? decisionFactors.map(factor => `
+            <div class="decision-row">
+              <header>
+                <strong>${escapeHtml(factor.label)}</strong>
+                <span class="chip">${Math.round((factor.contribution || 0) * 100)} % Beitrag</span>
+              </header>
+              <p class="muted">${escapeHtml(factor.entity_id || factor.factor_type)} · Zustand: ${escapeHtml(factor.state || "offen")} · Gewicht: ${Math.round((factor.weight || 0) * 100)} %</p>
+              <p>${(factor.evidence || []).map(escapeHtml).join(" ")}</p>
+            </div>
+          `).join("") : "<p class='muted'>Noch keine aktuelle Entscheidungsfaktoren berechnet.</p>"}
+        </div>
+      </details>
+    `;
+    const adaptivePanel = `
+      <details class="manual-context">
+        <summary>v1.2 Lernen, Rollback und Konflikte</summary>
+        <h3>Zeitprofile</h3>
+        <div class="metric-grid">
+          ${timeProfiles.length ? timeProfiles.map(profile => `
+            <div class="metric">
+              <strong>${escapeHtml(profile.label)}</strong>
+              ${escapeHtml(profile.sample_count)} Samples · ${escapeHtml(profile.dominant_state || "offen")}
+              <p class="muted">${Math.round((profile.confidence || 0) * 100)} % Profilklarheit</p>
+            </div>
+          `).join("") : "<div class='metric'><strong>Zeitprofile</strong>Noch keine Daten</div>"}
+        </div>
+        <h3>Modell-Snapshots</h3>
+        <div class="decision-list">
+          ${snapshots.length ? snapshots.slice(-5).reverse().map(snapshot => `
+            <div class="decision-row">
+              <header>
+                <strong>${escapeHtml(snapshot.version_id)}</strong>
+                <span class="chip">${snapshot.version_id === activeModelVersion ? "aktiv" : "Rollback möglich"}</span>
+              </header>
+              <p class="muted">${escapeHtml(snapshot.sample_count)} Samples · ${escapeHtml(snapshot.high_confidence_sample_count)} eindeutig · Ø ${Math.round((snapshot.average_confidence || 0) * 100)} %</p>
+              <p>${escapeHtml(snapshot.reason || "Kein Kommentar")}</p>
+              ${snapshot.version_id !== activeModelVersion ? `<button class="secondary compact" onclick="rollbackModel('${escapeHtml(record.actuator_entity_id)}', '${escapeHtml(snapshot.version_id)}')">Rollback</button>` : ""}
+            </div>
+          `).join("") : "<p class='muted'>Noch kein Modell-Snapshot gespeichert.</p>"}
+        </div>
+        <h3>Automatische Gewichtsanpassungen</h3>
+        <ul>${adaptiveUpdates.length ? adaptiveUpdates.slice(-8).reverse().map(update => `
+          <li><code>${escapeHtml(update.entity_id)}</code>: ${Math.round(update.previous_weight * 100)} % → ${Math.round(update.new_weight * 100)} %. ${escapeHtml(update.reason)}</li>
+        `).join("") : "<li>Noch keine automatische Gewichtsanpassung.</li>"}</ul>
+        <h3>Automation-Konflikte</h3>
+        <ul>${automationConflicts.length ? automationConflicts.map(conflict => `
+          <li><code>${escapeHtml(conflict.automation_entity_id)}</code>: <span class="${conflict.severity === "warning" ? "warn" : "muted"}">${escapeHtml(conflict.status)}</span> ${escapeHtml(conflict.reason)}</li>
+        `).join("") : "<li>Keine aktiven Automation-Konflikte erkannt.</li>"}</ul>
+      </details>
+    `;
    const learnedAutomationActions = record.behavior.patterns.filter(
      pattern => pattern.source === "automation",
    ).length;
@@ -978,6 +1113,9 @@ async function showActuator(actuatorId, evaluationMessage = "") {
        <button class="secondary" onclick="sendFeedback('${escapeHtml(record.actuator_entity_id)}', true)">Vorhersage korrekt</button>
        <button class="secondary" onclick="sendFeedback('${escapeHtml(record.actuator_entity_id)}', false)">Vorhersage falsch</button>
      </div>
+      ${safetyControls}
+      ${decisionArchive}
+      ${adaptivePanel}
      <h3>Passende Home-Assistant-Automationen</h3>
      <p class="muted">Bei einer Übernahme pausiert SillyHome diese Automationen. Beim Stoppen können sie gezielt fortgesetzt werden.</p>
      <button class="secondary compact" onclick="refreshRelatedAutomations('${escapeHtml(record.actuator_entity_id)}')">Automationen neu suchen</button>
@@ -1180,6 +1318,51 @@ async function sendFeedback(actuatorId, correct) {
  }
 }

+async function saveSafetyProfile(actuatorId) {
+  const confidence = Number(document.getElementById("safety-confidence")?.value || 82);
+  const cooldownRaw = document.getElementById("safety-cooldown")?.value || "";
+  const cooldown = cooldownRaw === "" ? null : Math.max(0, Number(cooldownRaw));
+  const profile = {
+    stage: document.getElementById("safety-stage")?.value || "shadow",
+    manual_block: Boolean(document.getElementById("safety-manual-block")?.checked),
+    min_confidence: Math.max(0, Math.min(100, Number.isFinite(confidence) ? confidence : 82)) / 100,
+    cooldown_seconds: Number.isFinite(cooldown) ? cooldown : null,
+    rules: [
+      {rule_id: "activation_ready", label: "Nur nach Lernfreigabe aktiv schalten", enabled: true, blocking: true, reason: "Der Aktor muss genug eindeutiges Verhalten gelernt haben."},
+      {rule_id: "confidence_threshold", label: "Mindest-Sicherheit einhalten", enabled: true, blocking: true, reason: "Vorhersagen unter der Schaltschwelle bleiben im Shadow-Modus."},
+      {rule_id: "cooldown", label: "Sicherheits-Cooldown gegen Hin-und-her-Schalten", enabled: true, blocking: true, reason: "Gleiche Zielzustände werden nicht zu schnell wiederholt."},
+      {rule_id: "manual_block", label: "Manuelle Sperre respektieren", enabled: true, blocking: true, reason: "Nutzer können jeden Aktor sofort blockieren."},
+    ],
+    note: "Sicherheitsprofil im Dashboard gespeichert",
+  };
+  try {
+    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/safety`, {
+      method: "POST",
+      body: JSON.stringify({safety: profile}),
+    });
+    invalidateDashboardCache();
+    await loadConfiguredActuators();
+    await showActuator(actuatorId, "Sicherheitsprofil gespeichert.");
+  } catch (error) {
+    alert(error.message);
+  }
+}
+
+async function rollbackModel(actuatorId, versionId) {
+  if (!confirm(`${actuatorId}: wirklich auf Modell ${versionId} zurückrollen?`)) return;
+  try {
+    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/model/rollback`, {
+      method: "POST",
+      body: JSON.stringify({version_id: versionId}),
+    });
+    invalidateDashboardCache();
+    await loadConfiguredActuators();
+    await showActuator(actuatorId, `Rollback auf ${versionId} ausgeführt.`);
+  } catch (error) {
+    alert(error.message);
+  }
+}
+
 async function setActivation(actuatorId, active, pauseMatchingAutomations, restorePausedAutomations) {
  const question = active
    ? pauseMatchingAutomations
--- a/docs/V1_0_0_OPERATING_GUIDE.md
+++ b/docs/V1_0_0_OPERATING_GUIDE.md
@@ -101,6 +101,21 @@ wget -qO /tmp/summary.json http://58adbe1e-sillyhome-next:8000/v1/actuators/summ
 wget -qO /tmp/dashboard.json http://58adbe1e-sillyhome-next:8000/v1/actuators/dashboard
 ```

+Wenn der Add-on-Container aus dem Agent-Host nicht direkt routbar ist, gilt der
+Home-Assistant-Supervisor als Verifikationsquelle:
+
+- Add-on-Info pruefen: Version, `version_latest`, `update_available`, `state`,
+  `boot` und `watchdog`.
+- Vor Updates eine Home-Assistant-Teil-Sicherung fuer **SillyHome Next**
+  erstellen.
+- Nach einem Store-Reload und Update muss `version == version_latest`,
+  `update_available == false`, `state == started`, `boot == auto` und
+  `watchdog == true` gelten.
+- Den HA-/Ingress-Tab nach jedem Update hart neu laden, weil Home Assistant
+  sonst alte HTML-/JavaScript-Ressourcen aus dem bestehenden Tab verwenden kann.
+- Rollback erfolgt ueber die vorherige Add-on-Teil-Sicherung oder den letzten
+  Git-Tag; beide Referenzen im Release-/Abnahmeprotokoll notieren.
+
 ## Rollback

 Der stabile Vor-1.0-Stand ist `v0.7.21`. Vor dem 1.0.0-Umbau wurde ein
--- a/docs/V1_0_ACCEPTANCE.md
+++ b/docs/V1_0_ACCEPTANCE.md
@@ -44,6 +44,12 @@ expliziter Freigabe.
  - `ruff check .`
  - `mypy app backend tests`
  - `git diff --check`
+- Performance-Budget:
+  - Automatisierter Test prueft Root-HTML und `/v1/actuators/dashboard` gegen
+    das 5-Sekunden-Budget mit kontrollierten Fake-HA-/Cache-Daten.
+- HA-/Ingress-Verifikation:
+  - Supervisor-Update, Add-on-Status, Watchdog, Backup, Ingress-Hard-Reload
+    und Rollback sind im Operating Guide dokumentiert.

 ## Teilweise Erfuellt

@@ -62,14 +68,10 @@ expliziter Freigabe.

 ## Offen Fuer v1.0.x

- Echte Dashboard-Performance-Budget-Tests, die Start-HTML und
-  `/v1/actuators/dashboard` gegen ein 5-Sekunden-Limit messen.
 - Dashboard-Jobstatus fuer Reconciliation, Training, Discovery und
  Automation-Refresh.
 - Mehr Entscheidungsstatistik pro Aktor: welche Sensoren wie stark
  beigetragen haben, wie sich Confidence und Sample Count entwickeln.
- Dokumentierte HA-Installationspruefung mit Supervisor-/Ingress-Hinweisen,
-  weil direkte Container-HTTP-Pruefung ausserhalb HA nicht immer routbar ist.

 ## Rollback

--- a/docs/V1_1_0_OPERATING_GUIDE.md
+++ b/docs/V1_1_0_OPERATING_GUIDE.md
@@ -0,0 +1,72 @@
+# SillyHome Next v1.1.0 Operating Guide
+
+## Ziel
+
+v1.1.0 macht das Dashboard zur Zentrale fuer Visualisierung, Einrichtung,
+Sicherheit und manuelles Gegensteuern. Autonomes Schalten bleibt ein kurzer
+lokaler Pfad: Vorhersage und Safety-Profil werden aus bereits vorhandenen Daten
+bewertet, danach folgt direkt der Home-Assistant-Serviceaufruf.
+
+## Sicherheitsmodell
+
+Jeder Aktor hat ein Safety-Profil:
+
+- `stage`: Beobachten, Vorschlagen, Shadow, Teilaktiv oder Aktiv.
+- `manual_block`: harte manuelle Sperre.
+- `min_confidence`: Mindest-Sicherheit fuer autonomes Schalten.
+- `cooldown_seconds`: optionaler Aktor-Cooldown gegen schnelles Hin-und-her.
+- Safety-Regeln: Freigabe, Confidence, Cooldown und manuelle Sperre.
+
+Ein Aktor schaltet nur, wenn alle lokalen Safety-Regeln frei sind, der
+Behavior-Modus aktiv ist, die Freigabe bereit ist, die Confidence passt, der
+Zielzustand noch nicht erreicht ist und der Cooldown abgelaufen ist.
+
+## Transparenz
+
+Die Aktor-Detailansicht trennt:
+
+- Wissen: belegte Fakten aus Historie, Zuordnung und Automationen.
+- Annahmen: heuristische Schluesse wie Zeit-/Kontext-Aehnlichkeit.
+- Unsicherheiten: geringe Datenmenge, unklare Quellen, Review-Bedarf oder
+  negatives Feedback.
+- Beitragsfaktoren: Sensoren, Kontextsignale, aktive Gewichtung und Beitrag.
+- Safety-Blocker: Gruende, warum nicht geschaltet wird.
+
+## Job-Queue
+
+Das Dashboard zeigt die letzten Jobs mit Status, Dauer, Fehler und
+Zusammenfassung. Sichtbar sind:
+
+- Discovery
+- Reconciliation
+- Training
+- Evaluation
+- Automation-Refresh
+
+Die Queue ist persistent in `job_queue.json` und dient als Betriebsanzeige. Sie
+blockiert nicht den Startpfad und nicht den Schaltpfad.
+
+## Manuelles Gegensteuern
+
+Im Dashboard koennen pro Aktor gesetzt werden:
+
+- manuelle Sicherheitssperre
+- Freigabestufe
+- Mindest-Confidence
+- optionaler Cooldown
+- Sensor-Gewichtungen und Gruppen-Gewichtungen
+- Kontextauswahl
+- Feedback: Vorhersage korrekt/falsch
+- HA-Automationen pausieren/fortsetzen
+
+## Qualitaetspruefung
+
+Vor Release:
+
+```bash
+.venv/bin/pytest -q
+.venv/bin/ruff check .
+.venv/bin/mypy app backend tests
+git diff --check
+node --check /tmp/sillyhome-dashboard.js
+```
--- a/docs/V1_2_0_OPERATING_GUIDE.md
+++ b/docs/V1_2_0_OPERATING_GUIDE.md
@@ -0,0 +1,62 @@
+# SillyHome Next v1.2.0 Operating Guide
+
+## Ziel
+
+v1.2.0 erweitert die sichere v1.1-Grundlage um adaptive Lernfunktionen. Diese
+Funktionen laufen bei Feedback, Training oder Automation-Refresh und blockieren
+nicht den direkten Schaltpfad.
+
+## Adaptive Gewichtung
+
+Feedback passt die Gewichtung aktuell beteiligter Kontextsignale vorsichtig an:
+
+- korrektes Feedback: +3 Prozentpunkte bis maximal 100 %
+- falsches Feedback: -8 Prozentpunkte bis minimal 10 %
+
+Die Aenderungen werden als `adaptive_weight_updates` gespeichert und im
+Dashboard angezeigt. Manuelle Gewichtungen bleiben weiter direkt korrigierbar.
+
+## Modell-Snapshots und Rollback
+
+Bei jedem Training wird ein Snapshot gespeichert:
+
+- Version-ID
+- Sample Count
+- eindeutig zugeordnete Handlungen
+- durchschnittliche Confidence
+- negative Feedbacks
+- Musterliste
+- Begruendung
+
+Ueber das Dashboard kann auf einen frueheren Snapshot zurueckgerollt werden.
+
+## Automation-Konflikte
+
+Beim Automation-Refresh markiert SillyHome Konflikte, wenn:
+
+- SillyHome fuer einen Aktor aktiv ist
+- eine passende Home-Assistant-Automation ebenfalls aktiv bleibt
+
+Pausierte Automationen werden als kontrolliert markiert.
+
+## Zeitprofile
+
+SillyHome bildet Profile fuer:
+
+- Nacht
+- Morgen
+- Tag
+- Abend
+- Wochenende
+
+Diese Profile zeigen Sample Count, dominanten Zielzustand und Profilklarheit.
+
+## Performance-Grenze
+
+v1.2-Funktionen duerfen den Schaltmoment nicht verlangsamen. Der direkte
+Schaltpfad bleibt:
+
+1. vorhandene aktuelle States nutzen
+2. lokale Safety-Pruefung
+3. direkter Home-Assistant-Serviceaufruf
+4. Persistenz der Entscheidung
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"

 [project]
 name = "sillyhome-next"
-version = "1.0.4"
+version = "1.2.0"
 description = "Lokales Smart-Home-Intelligenzsystem für Home Assistant"
 requires-python = ">=3.11"
 dependencies = [
--- a/tests/api/test_actuators.py
+++ b/tests/api/test_actuators.py
@@ -1,11 +1,13 @@
 from __future__ import annotations

+from time import perf_counter
 from datetime import datetime, timedelta
 from pathlib import Path

 from fastapi.testclient import TestClient

 from app.actuators.lifecycle import ActuatorReconciliationService
+from app.actuators.models import ModelSnapshot
 from app.actuators.store import ActuatorStore
 from app.behavior.engine import BehaviorEngine
 from app.config import Settings
@@ -29,6 +31,7 @@ class FakeHaReader(HaReader):
        self._entities = entities
        self._history = history
        self.read_entities_calls = 0
+        self.service_calls: list[tuple[str, str, dict[str, object]]] = []

    def read_entities(self) -> list[HaEntitySummary]:
        self.read_entities_calls += 1
@@ -85,6 +88,7 @@ class FakeHaReader(HaReader):
        service: str,
        service_data: dict[str, object],
    ) -> list[object]:
+        self.service_calls.append((domain, service, service_data))
        return []

    def find_automations_for_entity(
@@ -110,6 +114,7 @@ def _install_service(tmp_path: Path) -> None:
            unit_of_measurement="lx",
            friendly_name="Abstellkammer Helligkeit",
            area_name="Abstellkammer",
+            state="12",
        ),
        HaEntitySummary(
            entity_id="binary_sensor.abstellkammer_motion",
@@ -117,6 +122,7 @@ def _install_service(tmp_path: Path) -> None:
            device_class="motion",
            friendly_name="Abstellkammer Bewegung",
            area_name="Abstellkammer",
+            state="off",
        ),
        HaEntitySummary(
            entity_id="sensor.pfsense_interface_vpn_inbytes",
@@ -263,6 +269,88 @@ def test_weight_override_endpoint_updates_sensor_relevance(tmp_path: Path) -> No
    assert numeric["sensor.abstellkammer_illuminance"]["effective_weight"] == 0.75


+def test_safety_profile_can_block_actuator_manually(tmp_path: Path) -> None:
+    with TestClient(app) as client:
+        _install_service(tmp_path)
+        client.post("/v1/actuators", json={"actuator_entity_id": "light.abstellkammer"})
+
+        response = client.post(
+            "/v1/actuators/light.abstellkammer/safety",
+            json={
+                "safety": {
+                    "stage": "shadow",
+                    "manual_block": True,
+                    "min_confidence": 0.9,
+                    "cooldown_seconds": 120,
+                    "rules": [
+                        {
+                            "rule_id": "manual_block",
+                            "label": "Manuelle Sperre respektieren",
+                            "enabled": True,
+                            "blocking": True,
+                            "reason": "Test",
+                        }
+                    ],
+                    "note": "Test",
+                }
+            },
+        )
+
+    assert response.status_code == 200
+    payload = response.json()
+    assert payload["behavior"]["safety"]["manual_block"] is True
+    assert payload["behavior"]["safety"]["min_confidence"] == 0.9
+    assert payload["behavior"]["safety"]["cooldown_seconds"] == 120
+
+
+def test_feedback_adapts_sensor_weights_and_model_can_rollback(tmp_path: Path) -> None:
+    with TestClient(app) as client:
+        _install_service(tmp_path)
+        client.post(
+            "/v1/actuators",
+            json={"actuator_entity_id": "light.abstellkammer"},
+        )
+        record = app.state.actuator_store.get("light.abstellkammer")
+        version_id = "model-test"
+        snapshot = ModelSnapshot(
+            version_id=version_id,
+            sample_count=1,
+            high_confidence_sample_count=1,
+            average_confidence=0.9,
+            patterns=[],
+            reason="Test-Snapshot",
+        )
+        app.state.actuator_store.upsert(
+            record.model_copy(
+                update={
+                    "behavior": record.behavior.model_copy(
+                        update={
+                            "model_snapshots": [snapshot],
+                            "active_model_version": "model-current",
+                            "sample_count": 2,
+                        }
+                    )
+                }
+            )
+        )
+
+        feedback = client.post(
+            "/v1/actuators/light.abstellkammer/feedback",
+            json={"correct": False, "expected_state": "off"},
+        )
+        rollback = client.post(
+            "/v1/actuators/light.abstellkammer/model/rollback",
+            json={"version_id": version_id},
+        )
+
+    assert feedback.status_code == 200
+    feedback_payload = feedback.json()
+    assert feedback_payload["behavior"]["adaptive_weight_updates"]
+    assert feedback_payload["manual_override"]["sensor_weights"]
+    assert rollback.status_code == 200
+    assert rollback.json()["behavior"]["active_model_version"] == version_id
+
+
 def test_summary_is_lightweight_and_uses_cached_entity_metadata(tmp_path: Path) -> None:
    with TestClient(app) as client:
        _install_service(tmp_path)
@@ -297,6 +385,46 @@ def test_dashboard_overview_uses_cache_without_ha_roundtrip(tmp_path: Path) -> N
    assert payload["cache"]["entity_count"] == 4
    assert payload["actuators"][0]["friendly_name"] == "Abstellkammer Licht"
    assert payload["discovery_groups"]
+    assert payload["jobs"]["jobs"][-1]["kind"] == "discovery"
+
+
+def test_reconciliation_run_records_visible_job_queue(tmp_path: Path) -> None:
+    with TestClient(app) as client:
+        _install_service(tmp_path)
+        client.post("/v1/actuators", json={"actuator_entity_id": "light.abstellkammer"})
+
+        response = client.post("/v1/actuators/reconciliation/run")
+        jobs = client.get("/v1/actuators/job-queue/state")
+
+    assert response.status_code == 200
+    assert jobs.status_code == 200
+    payload = jobs.json()
+    assert [job["kind"] for job in payload["jobs"][-3:]] == [
+        "reconciliation",
+        "training",
+        "evaluation",
+    ]
+    assert payload["jobs"][-1]["status"] == "completed"
+
+
+def test_dashboard_start_path_stays_within_five_second_budget(tmp_path: Path) -> None:
+    with TestClient(app) as client:
+        _install_service(tmp_path)
+        client.get("/v1/actuators/discovery")
+        client.post("/v1/actuators", json={"actuator_entity_id": "light.abstellkammer"})
+
+        root_started_at = perf_counter()
+        root_response = client.get("/")
+        root_elapsed = perf_counter() - root_started_at
+
+        dashboard_started_at = perf_counter()
+        dashboard_response = client.get("/v1/actuators/dashboard")
+        dashboard_elapsed = perf_counter() - dashboard_started_at
+
+    assert root_response.status_code == 200
+    assert dashboard_response.status_code == 200
+    assert root_elapsed < 5.0
+    assert dashboard_elapsed < 5.0


 def test_discovery_reads_entities_once_and_reuses_them(tmp_path: Path) -> None:
Author	SHA1	Message	Date
Otto	2ec2c64cba	Add adaptive learning and model rollback Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-17 18:41:03 +02:00
Otto	0101596e93	Add safety dashboard and decision transparency Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-17 18:26:49 +02:00
Otto	ca253d1e6c	Fix dashboard text overflow and close v1 docs gaps Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-17 11:53:25 +02:00