diff --git a/src/evaluation/metrics.py b/src/evaluation/metrics.py
index 2293c82..0180ca1 100644
--- a/src/evaluation/metrics.py
+++ b/src/evaluation/metrics.py
@@ -33,13 +33,17 @@ def expected_calibration_error(y_true, y_prob, n_bins: int = 10) -> float:
 
 def binary_metrics(y_true, y_prob, threshold: float = 0.5) -> dict:
     y_pred = (np.asarray(y_prob) >= threshold).astype(int)
+    precision = float(precision_score(y_true, y_pred, zero_division=0))
+    recall = float(recall_score(y_true, y_pred, zero_division=0))
+    f2 = (5.0 * precision * recall) / (4.0 * precision + recall + 1e-12)
+
     return {
         "roc_auc": safe_auc(y_true, y_prob, "roc"),
         "pr_auc": safe_auc(y_true, y_prob, "pr"),
         "f1": float(f1_score(y_true, y_pred, zero_division=0)),
-        "f2": float((5 * precision_score(y_true, y_pred, zero_division=0) * recall_score(y_true, y_pred, zero_division=0)) / (4 * precision_score(y_true, y_pred, zero_division=0) + recall_score(y_true, y_pred, zero_division=0) + 1e-12)),
-        "precision": float(precision_score(y_true, y_pred, zero_division=0)),
-        "recall": float(recall_score(y_true, y_pred, zero_division=0)),
+        "f2": float(f2),
+        "precision": precision,
+        "recall": recall,
         "brier": float(brier_score_loss(y_true, y_prob)),
         "ece": expected_calibration_error(y_true, y_prob),
     }
diff --git a/tests/test_binary_metrics_polish.py b/tests/test_binary_metrics_polish.py
new file mode 100644
index 0000000..b5cf7d7
--- /dev/null
+++ b/tests/test_binary_metrics_polish.py
@@ -0,0 +1,19 @@
+"""Regression checks compatible with the project's lightweight test runner."""
+
+from src.evaluation.metrics import binary_metrics
+
+
+def test_f2_uses_reported_precision_and_recall() -> None:
+    cases = [
+        ([0, 1, 1, 0], [0.1, 0.9, 0.4, 0.3], 1.0, 0.5),
+        ([0, 0, 1, 1], [0.8, 0.7, 0.2, 0.1], 0.0, 0.0),
+    ]
+    for labels, scores, expected_precision, expected_recall in cases:
+        result = binary_metrics(labels, scores, threshold=0.5)
+        assert abs(result["precision"] - expected_precision) < 1e-12
+        assert abs(result["recall"] - expected_recall) < 1e-12
+        expected_f2 = (
+            5.0 * expected_precision * expected_recall
+            / (4.0 * expected_precision + expected_recall + 1e-12)
+        )
+        assert abs(result["f2"] - expected_f2) < 1e-12