diff --git a/python/fast_mlsirm/llm_judge.py b/python/fast_mlsirm/llm_judge.py
index 61d2934f7..cc02e2395 100644
--- a/python/fast_mlsirm/llm_judge.py
+++ b/python/fast_mlsirm/llm_judge.py
@@ -263,7 +263,7 @@ def _criteria(values: Iterable[JudgeCriterion | Mapping[str, Any]]) -> tuple[Jud
                 weight=value.get("weight", 1.0),
             )
         else:
-            raise TypeError("criteria must contain JudgeCriterion or mapping values")
+            raise ValueError("criteria must contain JudgeCriterion or mapping values")
         normalized.append(criterion)
     if not 1 <= len(normalized) <= MAX_JUDGE_CRITERIA:
         raise ValueError(f"criteria must contain 1..{MAX_JUDGE_CRITERIA} values")
diff --git a/tests/test_llm_judge.py b/tests/test_llm_judge.py
index 8e1244a47..b7359442e 100644
--- a/tests/test_llm_judge.py
+++ b/tests/test_llm_judge.py
@@ -269,6 +269,16 @@ def __float__(self):
     assert _HookedFloat.invoked is False
 
 
+def test_judge_criteria_reject_non_contract_values_with_value_error() -> None:
+    """Arbitrary criterion elements must fail through the stable benign error contract."""
+    with pytest.raises(ValueError, match="JudgeCriterion or mapping"):
+        ContextualOrchestratorJudge(_FakeOrchestrator(_payload())).judge(
+            task="task",
+            answer="answer",
+            criteria=[object()],
+        )
+
+
 if __name__ == "__main__":
     test_judge_uses_contextual_orchestrator_route_and_reports_usage()
     test_judge_rejects_malformed_decisions_and_derives_acceptance()
@@ -279,4 +289,5 @@ def __float__(self):
     test_category_judgment_rejects_non_integral_categories()
     test_judge_rejects_missing_or_malformed_model_fields()
     test_judge_criteria_reject_invalid_runtime_types()
+    test_judge_criteria_reject_non_contract_values_with_value_error()
     print("ok")