diff --git a/seapig/scores/knn.py b/seapig/scores/knn.py
index bee843c..c4a2344 100644
--- a/seapig/scores/knn.py
+++ b/seapig/scores/knn.py
@@ -34,7 +34,7 @@ class KNNScore(EmbeddingScore, FAISSIndexMixin, ABC):
     ----------
     k : int, default 1
         Number of nearest neighbors used to compute the distance score.
-    stat : {'max', 'mean', 'median', 'min'}, default 'max'
+    stat : {'max', 'mean', 'median', 'min'}, default 'min'
         Statistic applied to aggregate distances across the k neighbors.
     pca : `TensorPCA` or None, default None
         Optional `TensorPCA` object for dimensionality reduction prior to scoring.
@@ -57,7 +57,7 @@ class KNNScore(EmbeddingScore, FAISSIndexMixin, ABC):
     def __init__(
         self,
         k: int = 1,
-        stat: str = "max",
+        stat: str = "min",
         pca: TensorPCA | None = None,
         save_index: bool | Path = False,
     ) -> None:
@@ -110,7 +110,7 @@ def _filter_outliers(self, q: float | bool = False) -> None:
         if self.index is None:
             self._setup_index()  # temporary index
         scores, _ = self._distance(self.ref_embeddings, offset=1)
-        scores = self._stat(-scores)
+        scores = self._stat(scores)
         threshold = torch.quantile(scores.float(), q=q)
         index = scores >= threshold
         self.ref_embeddings = self.ref_embeddings[index, :]
@@ -210,7 +210,7 @@ class EuclideanScore(KNNScore):
     ----------
     k : int, default 1
         Number of nearest neighbors to use.
-    stat : {'max', 'mean', 'median', 'min'}, default 'max'
+    stat : {'max', 'mean', 'median', 'min'}, default 'min'
         Statistic to aggregate distances across the k neighbors.
     pca : `TensorPCA` or None, default None
         Optional `TensorPCA` object for dimensionality reduction prior to scoring.
@@ -241,7 +241,7 @@ class EuclideanScore(KNNScore):
     def __init__(
         self,
         k: int = 1,
-        stat: str = "max",
+        stat: str = "min",
         pca: TensorPCA | None = None,
         save_index: bool | Path = False,
     ) -> None:
@@ -287,7 +287,7 @@ class CosineScore(KNNScore):
     ----------
     k : int, default 1
         Number of nearest neighbors to use.
-    stat : {'max', 'mean', 'median', 'min'}, default 'max'
+    stat : {'max', 'mean', 'median', 'min'}, default 'min'
         Statistic to aggregate distances across the k neighbors.
     pca : `TensorPCA` or None, default None
         Optional `TensorPCA` object for dimensionality reduction prior to scoring.
@@ -307,7 +307,7 @@ class CosineScore(KNNScore):
     def __init__(
         self,
         k: int = 1,
-        stat: str = "max",
+        stat: str = "min",
         pca: TensorPCA | None = None,
         save_index: bool | Path = False,
     ) -> None:
@@ -366,7 +366,7 @@ class MahalanobisScore(KNNScore):
     ----------
     k : int, default 1
         Number of nearest neighbors to use.
-    stat : {'max', 'mean', 'median', 'min'}, default 'max'
+    stat : {'max', 'mean', 'median', 'min'}, default 'min'
         Statistic to aggregate distances across the k neighbors.
     pca : `TensorPCA` or None, default None
         Optional `TensorPCA` object for dimensionality reduction prior to scoring.
@@ -387,7 +387,7 @@ class MahalanobisScore(KNNScore):
     def __init__(
         self,
         k: int = 1,
-        stat: str = "max",
+        stat: str = "min",
         pca: TensorPCA | None = None,
         save_index: bool | Path = False,
     ) -> None:
diff --git a/tests/test_knn.py b/tests/test_knn.py
index f9531bf..500500a 100644
--- a/tests/test_knn.py
+++ b/tests/test_knn.py
@@ -323,3 +323,18 @@ def test_knnstat_all_branches(
     result = score._stat(dists)
     expected_val = expected(dists)
     torch.testing.assert_close(result[0], expected_val)
+
+
+def test_k_returns_diff():
+    ref = torch.rand(5000, 256)
+    query = torch.rand(100, 256)
+
+    s1 = EuclideanScore(k=1)
+    s2 = EuclideanScore(k=10)
+    s1.fit(ref)
+    s2.fit(ref)
+
+    s_1 = s1.score(query)
+    s_2 = s2.score(query)
+
+    assert not torch.allclose(s_1, s_2)