diff --git a/app/processors/video_processor.py b/app/processors/video_processor.py
index e0f367fa..61a5f226 100644
--- a/app/processors/video_processor.py
+++ b/app/processors/video_processor.py
@@ -1969,6 +1969,9 @@ def stop_processing(self) -> bool:
         self.recording = False
         self.triggered_by_job_manager = False
         self.active_output_folder = ""
+        # Disarm the batch recognition flag to prevent UI side-effects after an abort
+        if hasattr(self.main_window, "force_recognition_in_batch"):
+            self.main_window.force_recognition_in_batch = False
         self._cancel_single_frame_preview_state()
 
         # 2. Stop utility timers and audio
diff --git a/app/ui/widgets/actions/job_manager_actions.py b/app/ui/widgets/actions/job_manager_actions.py
index fe56a936..c96767bb 100644
--- a/app/ui/widgets/actions/job_manager_actions.py
+++ b/app/ui/widgets/actions/job_manager_actions.py
@@ -1603,6 +1603,12 @@ def load_job_settings(main_window: "MainWindow", job_data: dict):
         # 4. Load markers.
         _load_job_markers(main_window, job_data)
 
+        # --- Arm the recognition flag for the new job ---
+        # We MUST set this to True here in the Main Thread before processing starts.
+        # This forces the sequential_detector's feeder_thread to run ArcFace on the
+        # first frame, preventing identity bleed (Fast-Path bypass) from the previous job.
+        main_window.force_recognition_in_batch = True
+
         # Calculate assigned_input_embedding here for KV injection
         for face_id, target_face_button in main_window.target_faces.items():
             print(