diff --git a/promptpulse/data.py b/promptpulse/data.py
index c7f1c05..dd624d0 100644
--- a/promptpulse/data.py
+++ b/promptpulse/data.py
@@ -25,10 +25,17 @@ def load_dataset(path: str | Path = DEFAULT_DATASET) -> list[dict[str, Any]]:
         missing = required - row.keys()
         if missing:
             raise ValueError(f"Dataset row {index} is missing: {sorted(missing)}")
+        for field in required:
+            if not isinstance(row[field], str) or not row[field].strip():
+                raise ValueError(f"Dataset row {index} requires a non-empty string for {field}.")
         if row["id"] in seen:
             raise ValueError(f"Duplicate dataset id: {row['id']}")
         seen.add(row["id"])
         row.setdefault("required_terms", [])
         row.setdefault("forbidden_terms", [])
+        for field in ("required_terms", "forbidden_terms"):
+            terms = row[field]
+            if not isinstance(terms, list) or any(not isinstance(term, str) or not term.strip() for term in terms):
+                raise ValueError(f"Dataset row {index} requires a list of non-empty strings for {field}.")
 
     return data
diff --git a/tests/test_evaluation.py b/tests/test_evaluation.py
index fd9601d..98232ea 100644
--- a/tests/test_evaluation.py
+++ b/tests/test_evaluation.py
@@ -1,5 +1,7 @@
 from promptpulse.data import load_dataset
 from promptpulse.evaluation import evaluate_response, policy_compliance
+import json
+import pytest
 
 
 def test_dataset_schema_and_unique_ids():
@@ -8,6 +10,22 @@ def test_dataset_schema_and_unique_ids():
     assert len({row["id"] for row in rows}) == len(rows)
 
 
+def test_malformed_policy_terms_cannot_be_scored_as_character_checks(tmp_path):
+    row = dict(load_dataset()[0], required_terms="30 days")
+    path = tmp_path / "invalid.json"
+    path.write_text(json.dumps([row]))
+    with pytest.raises(ValueError, match="required_terms"):
+        load_dataset(path)
+
+
+def test_empty_reference_is_rejected_before_evaluation(tmp_path):
+    row = dict(load_dataset()[0], reference_context=" ")
+    path = tmp_path / "invalid.json"
+    path.write_text(json.dumps([row]))
+    with pytest.raises(ValueError, match="reference_context"):
+        load_dataset(path)
+
+
 def test_expected_answers_clear_deterministic_gate():
     for row in load_dataset():
         result = evaluate_response(