Gray out Train Aligners until a dataset has manual alignments

config/profiles/explanatory/pipeline_definitions.yaml

@@ -38,7 +38,7 @@ pipeline:
 
       Training improves alignment accuracy by adapting models to your speakers, recording conditions, or dialect. Requires a dataset that already has alignments (from a pretrained model) to use as training data.
 
-      *Skip this step if using pretrained models is sufficient for your needs.*
+      *Train Aligners is unavailable in the left sidebar until a dataset with manual alignments is registered.*
 
       ---
 

src/voxkit/gui/__init__.py

@@ -158,6 +158,10 @@ def build_feedback_mailto_url(
     QListWidget::item:hover {
         background-color: #b0cef2;
     }
+    QListWidget::item:disabled {
+        color: #a8a8a8;
+        background-color: transparent;
+    }
     QWidget#centralWidget {
         background-color: #f5f7fa;
     }

src/voxkit/gui/pages/pipeline/__init__.py

@@ -84,6 +84,7 @@ def on_settings(self):
     QHBoxLayout,
     QLabel,
     QListWidget,
+    QListWidgetItem,
     QPushButton,
     QScrollArea,
     QVBoxLayout,
@@ -92,6 +93,7 @@ def on_settings(self):
 
 from voxkit.gui.components import AnimatedStackedWidget
 from voxkit.gui.styles import Buttons, Containers, Labels
+from voxkit.storage import datasets
 
 if TYPE_CHECKING:
     from voxkit.config.pipeline_config import PipelineConfig
@@ -164,13 +166,18 @@ def init_ui(self):
         # Store mapping of menu index to stacker for reload
         self.stacker_instances = []
 
+        # Store menu items by step id so availability can be toggled later
+        self._menu_items: dict[str, QListWidgetItem] = {}
+
         # Right side - Stacked widget for different pipeline pages
         self.stacked_widget = AnimatedStackedWidget()
 
         # Dynamically create menu items and stackers from configuration
         for step in self.config.enabled_steps:
             # Add menu item
-            self.menu_list.addItem(step.label)
+            menu_item = QListWidgetItem(step.label)
+            self.menu_list.addItem(menu_item)
+            self._menu_items[step.id] = menu_item
 
             # Get the stacker class from registry
             stacker_class = STACKER_REGISTRY.get(step.stacker_class)
@@ -264,6 +271,8 @@ def toggle():
         self.menu_list.currentRowChanged.connect(self.change_page)
         self.menu_list.setCurrentRow(0)
 
+        self._refresh_training_availability()
+
     def reload(self):
         """Reload models and datasets in the pipeline pages.
 
@@ -303,6 +312,32 @@ def reload(self):
                 if hasattr(stacker_widget, "reload_datasets"):
                     stacker_widget.reload_datasets()
 
+        self._refresh_training_availability()
+
+    def _refresh_training_availability(self):
+        """Gray out the Train Aligners step unless a dataset has a manual alignment.
+
+        Training only produces a meaningful model when it can learn from a
+        hand-corrected alignment, so this step is disabled until at least one
+        dataset qualifies. Why it's disabled is explained in the Ⓑ Train
+        Aligners section of the Pipeline Overview, not in the sidebar itself
+        (too narrow for an inline explanation).
+        """
+        menu_item = self._menu_items.get("training")
+        stacker_widget = next(
+            (widget for step_id, _, widget in self.stacker_instances if step_id == "training"),
+            None,
+        )
+        if menu_item is None or stacker_widget is None:
+            return
+
+        if datasets.any_dataset_has_manual_alignments():
+            menu_item.setFlags(menu_item.flags() | Qt.ItemFlag.ItemIsEnabled)
+            stacker_widget.setEnabled(True)
+        else:
+            menu_item.setFlags(menu_item.flags() & ~Qt.ItemFlag.ItemIsEnabled)
+            stacker_widget.setEnabled(False)
+
     def change_page(self, index):
         """Change the displayed page based on menu selection with animation"""
         if index >= 0:  # Valid index

src/voxkit/gui/pages/pipeline/training_stacker.py

@@ -297,6 +297,15 @@ def reload_datasets(self):
 
     def build_ui(self):
         """Build the training UI."""
+        note_label = QLabel(
+            "Note: training only makes sense for datasets with manual (hand-corrected) "
+            "alignments — models trained on machine-generated alignments alone tend to "
+            "just reinforce whatever errors are already there."
+        )
+        note_label.setWordWrap(True)
+        note_label.setStyleSheet(Labels.INFO)
+        self.content_layout.addWidget(note_label)
+
         # Model Selection Panel
         engines_dict = {
             engine_id: engine

src/voxkit/storage/datasets.py

@@ -20,6 +20,7 @@
 - **create_dataset**: Create a new dataset with metadata and directories
 - **get_dataset_metadata**: Retrieve metadata for a specific dataset
 - **list_datasets_metadata**: List all existing datasets
+- **any_dataset_has_manual_alignments**: Whether any dataset has a hand alignment
 - **update_dataset_metadata**: Update metadata fields for a specific dataset
 - **delete_dataset**: Delete a registered dataset and its metadata
 - **export_dataset**: Export a dataset to a specified output path
@@ -321,6 +322,15 @@ def list_datasets_metadata() -> List[DatasetMetadata]:
         return []
 
 
+def any_dataset_has_manual_alignments() -> bool:
+    """Whether at least one registered dataset has a hand/manual alignment.
+
+    Training only produces a meaningful model when it can learn from a
+    manually-corrected alignment, so this gates the Train Aligners step.
+    """
+    return any(d.get("hand_alignments_path") for d in list_datasets_metadata())
+
+
 def update_dataset_metadata(
     dataset_id: str,
     updates: dict,