diff --git a/slime/rollout/rm_hub/math_dapo_utils.py b/slime/rollout/rm_hub/math_dapo_utils.py
index 915c3aa480..1cab28ca0a 100644
--- a/slime/rollout/rm_hub/math_dapo_utils.py
+++ b/slime/rollout/rm_hub/math_dapo_utils.py
@@ -15,6 +15,7 @@
 
 import re
 import signal
+from decimal import Decimal, InvalidOperation
 
 
 def last_boxed_only_string(string: str) -> str | None:
@@ -63,7 +64,6 @@ def remove_boxed(s: str) -> str:
 
 
 class timeout:
-
     def __init__(self, seconds=1, error_message="Timeout"):
         self.seconds = seconds
         self.error_message = error_message
@@ -207,7 +207,15 @@ def is_correct_minerva(
     else:
         gt = normalize_final_answer(gt)
 
-    gt = str(int(float(gt)))  # in dapo, all answers are integers
+    try:
+        numeric_gt = Decimal(gt)
+    except InvalidOperation as exc:
+        raise ValueError(f"DAPO ground-truth label must be an integer, got {gt!r}") from exc
+
+    if not numeric_gt.is_finite() or numeric_gt != numeric_gt.to_integral_value():
+        raise ValueError(f"DAPO ground-truth label must be an integer, got {gt!r}")
+
+    gt = str(int(numeric_gt))
 
     return (pred == gt), pred
 
diff --git a/tests/test_rm_math_dapo.py b/tests/test_rm_math_dapo.py
index 3a5e9a28b9..9d241b34ce 100644
--- a/tests/test_rm_math_dapo.py
+++ b/tests/test_rm_math_dapo.py
@@ -153,14 +153,37 @@ def test_is_correct_strict_box_no_box_returns_minus_one():
 
 @pytest.mark.unit
 def test_is_correct_minerva_matches_int_answer():
-    """Minerva path expects gt to coerce via ``int(float(gt))`` (line 210),
-    so floats / int-strings collapse to canonical int strings before
-    comparison."""
     correct, pred = is_correct_minerva("Long solution. Answer: 42", "42")
     assert correct is True
     assert pred == "42"
 
 
+@pytest.mark.unit
+@pytest.mark.parametrize("ground_truth", ["42.0", "4.2e1"])
+def test_is_correct_minerva_accepts_integral_numeric_labels(ground_truth):
+    correct, pred = is_correct_minerva("Answer: 42", ground_truth)
+    assert correct is True
+    assert pred == "42"
+
+
+@pytest.mark.unit
+def test_is_correct_minerva_preserves_large_integer_precision():
+    ground_truth = "9007199254740993"
+    correct, pred = is_correct_minerva(f"Answer: {ground_truth}", ground_truth)
+    assert correct is True
+    assert pred == ground_truth
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("ground_truth", ["0.5", "1/2", "1+2", "NaN", "Infinity"])
+def test_is_correct_minerva_rejects_non_integer_labels(ground_truth):
+    with pytest.raises(
+        ValueError,
+        match="DAPO ground-truth label must be an integer",
+    ):
+        is_correct_minerva("Answer: 0", ground_truth)
+
+
 @pytest.mark.unit
 def test_is_correct_minerva_takes_last_answer_match():
     """Multiple ``Answer:`` lines → ``re.findall`` returns the list and