diff --git a/perceptionmetrics/evaluation/__init__.py b/perceptionmetrics/evaluation/__init__.py
new file mode 100644
index 00000000..4258807c
--- /dev/null
+++ b/perceptionmetrics/evaluation/__init__.py
@@ -0,0 +1,3 @@
+from perceptionmetrics.evaluation.segmentation_evaluator import SegmentationEvaluator
+
+__all__ = ["SegmentationEvaluator"]
diff --git a/perceptionmetrics/evaluation/segmentation_evaluator.py b/perceptionmetrics/evaluation/segmentation_evaluator.py
new file mode 100644
index 00000000..66035fab
--- /dev/null
+++ b/perceptionmetrics/evaluation/segmentation_evaluator.py
@@ -0,0 +1,60 @@
+from typing import Any, List
+
+import numpy as np
+from PIL import Image
+
+import perceptionmetrics.utils.segmentation_metrics as um
+
+
+class SegmentationEvaluator:
+    """Minimal prototype evaluator for segmentation models.
+
+    This class extracts only the common evaluation loop:
+    dataset iteration, `model.predict()` invocation, and metric aggregation.
+
+    The current prototype intentionally excludes ontology translation, prediction
+    saving, per-sample reports, batching, and backend-specific behavior so it can
+    be integrated safely and extended incrementally in the future.
+
+    Note:
+        This implementation is intentionally minimal and mirrors only the
+        default evaluation path to ensure backward compatibility.
+    """
+
+    def __init__(self, n_classes: int):
+        """Initialize metric aggregation.
+
+        :param n_classes: Number of segmentation classes.
+        :type n_classes: int
+        """
+        self.n_classes = n_classes
+        self.metrics_factory = um.SegmentationMetricsFactory(n_classes)
+
+    def evaluate(self, model: Any, dataset: Any, splits: List[str]):
+        """Evaluate a segmentation model over dataset samples.
+
+        :param model: Segmentation model exposing `predict(image)`.
+        :type model: Any
+        :param dataset: Dataset object exposing `dataset` and `make_fname_global`.
+        :type dataset: Any
+        :param splits: Dataset splits to evaluate.
+        :type splits: List[str]
+        :return: Aggregated segmentation metrics factory.
+        :rtype: um.SegmentationMetricsFactory
+        """
+        if "split" not in dataset.dataset.columns:
+            raise ValueError("Dataset must contain 'split' column")
+
+        dataset.make_fname_global()
+        df = dataset.dataset[dataset.dataset["split"].isin(splits)].copy()
+
+        for _, row in df.iterrows():
+            image = Image.open(row["image"]).convert("RGB")
+            label = np.array(Image.open(row["label"]))
+            pred = model.predict(image)
+
+            pred = np.array(pred)
+
+            self.metrics_factory.update(pred, label, None)
+
+        return self.metrics_factory
diff --git a/perceptionmetrics/models/torch_segmentation.py b/perceptionmetrics/models/torch_segmentation.py
index 5b71ff78..f935ec33 100644
--- a/perceptionmetrics/models/torch_segmentation.py
+++ b/perceptionmetrics/models/torch_segmentation.py
@@ -389,6 +389,26 @@ def eval(
         if predictions_outdir is not None:
             os.makedirs(predictions_outdir, exist_ok=True)
 
+        splits = [split] if isinstance(split, str) else split
+        ignored_classes = self.model_cfg.get("ignored_classes", [])
+        use_evaluator = (
+            ontology_translation is None
+            and predictions_outdir is None
+            and not results_per_sample
+            and len(ignored_classes) == 0
+        )
+
+        if use_evaluator:
+            # Use minimal evaluator only for default/simple path to avoid
+            # modifying advanced evaluation behavior (ontology, saving, etc.).
+            from perceptionmetrics.evaluation.segmentation_evaluator import (
+                SegmentationEvaluator,
+            )
+
+            evaluator = SegmentationEvaluator(n_classes=len(self.ontology))
+            metrics_factory = evaluator.evaluate(self, dataset, splits)
+            return um.get_metrics_dataframe(metrics_factory, self.ontology)
+
         # Build a LUT for transforming ontology if needed (aligned with TorchLiDARSegmentationModel.eval)
         eval_ontology = self.ontology
 
@@ -420,7 +440,7 @@ def eval(
             dataset,
             transform=self.transform_input,
             target_transform=self.transform_label,
-            splits=[split] if isinstance(split, str) else split,
+            splits=splits,
         )
 
         dataloader = DataLoader(