diff --git a/src/rtichoke/performance_data/performance_data.py b/src/rtichoke/performance_data/performance_data.py
index d734ac99..27bac46f 100644
--- a/src/rtichoke/performance_data/performance_data.py
+++ b/src/rtichoke/performance_data/performance_data.py
@@ -19,6 +19,31 @@
 import numpy as np
 
 
+_PERFORMANCE_DATA_COLUMNS = [
+    "reference_group",
+    "stratified_by",
+    "chosen_cutoff",
+    "true_positives",
+    "true_negatives",
+    "false_positives",
+    "false_negatives",
+    "predicted_positives",
+    "predicted_negatives",
+    "real_positives",
+    "real_negatives",
+    "n",
+    "sensitivity",
+    "specificity",
+    "ppv",
+    "npv",
+    "false_positive_rate",
+    "lift",
+    "net_benefit",
+    "net_benefit_interventions_avoided",
+    "ppcr",
+]
+
+
 def _probs_with_r_binary_cutoff_semantics(
     probs: Dict[str, np.ndarray], by: float
 ) -> Dict[str, np.ndarray]:
@@ -192,6 +217,6 @@ def prepare_performance_data(
 
     performance_data = _turn_cumulative_aj_to_performance_data(cumulative_aj_data)
 
-    return performance_data.sort(
+    return performance_data.select(_PERFORMANCE_DATA_COLUMNS).sort(
         ["reference_group", "stratified_by", "chosen_cutoff"]
     )
diff --git a/tests/test_performance_schema_order.py b/tests/test_performance_schema_order.py
new file mode 100644
index 00000000..5a21bc9d
--- /dev/null
+++ b/tests/test_performance_schema_order.py
@@ -0,0 +1,72 @@
+import numpy as np
+import polars as pl
+from polars.testing import assert_frame_equal
+
+from rtichoke import prepare_performance_data
+
+
+PROBS = {"model": np.array([0.1, 0.2, 0.5, 0.5, 0.8, 0.9])}
+REALS = np.array([0, 0, 1, 0, 1, 1])
+
+
+EXPECTED_COLUMNS = [
+    "reference_group",
+    "stratified_by",
+    "chosen_cutoff",
+    "true_positives",
+    "true_negatives",
+    "false_positives",
+    "false_negatives",
+    "predicted_positives",
+    "predicted_negatives",
+    "real_positives",
+    "real_negatives",
+    "n",
+    "sensitivity",
+    "specificity",
+    "ppv",
+    "npv",
+    "false_positive_rate",
+    "lift",
+    "net_benefit",
+    "net_benefit_interventions_avoided",
+    "ppcr",
+]
+
+
+def test_binary_performance_schema_order_is_stable_across_stratification():
+    for stratified_by in [
+        ("probability_threshold",),
+        ("ppcr",),
+        ("probability_threshold", "ppcr"),
+    ]:
+        result = prepare_performance_data(
+            probs=PROBS,
+            reals=REALS,
+            stratified_by=stratified_by,
+            by=0.5,
+        )
+
+        assert result.columns == EXPECTED_COLUMNS
+
+
+def test_combined_ppcr_values_match_ppcr_only():
+    ppcr_only = prepare_performance_data(
+        probs=PROBS,
+        reals=REALS,
+        stratified_by=("ppcr",),
+        by=0.5,
+    ).sort(["reference_group", "chosen_cutoff"])
+
+    combined_ppcr = (
+        prepare_performance_data(
+            probs=PROBS,
+            reals=REALS,
+            stratified_by=("probability_threshold", "ppcr"),
+            by=0.5,
+        )
+        .filter(pl.col("stratified_by") == "ppcr")
+        .sort(["reference_group", "chosen_cutoff"])
+    )
+
+    assert_frame_equal(combined_ppcr, ppcr_only)