Sync from GitHub via hub-sync

learning_engine.py

@@ -56,6 +56,16 @@ def next_card(session: Session) -> Card | None:
 # ---- Grading ---------------------------------------------------------------
 
 def grade_answer(card: Card, user_answer: str) -> GradeResult:
+    # An empty answer is unambiguously a miss — short-circuit before any model
+    # call. (The grader otherwise ignores the blank input and "grades" the
+    # reference answer itself, hallucinating a 4/5 "correct".) Also saves a call.
+    if not (user_answer or "").strip():
+        return new_grade(
+            0,
+            f"No answer given. The reference answer is: {card['answer']}",
+            card["topic"],
+        )
+
     if llm.STUB:
         # Trivial heuristic so the stub demo "feels" responsive.
         ans = (user_answer or "").strip().lower()

test_grade_answer.py

@@ -88,9 +88,22 @@ def test_out_of_range_score_rejected():
     print("ok  out-of-range score rejected -> safe default")
 
 
+def test_empty_answer_short_circuits_to_zero():
+    # An empty answer is a miss — score 0 with no model call (the model otherwise
+    # ignores the blank input and hallucinates a 4/5 "correct").
+    llm.chat, calls = _fake_chat(['{"score": 5, "explanation": "x"}'])
+    for blank in ("", "   ", "\n\t"):
+        g = le.grade_answer(_card(), blank)
+        assert g["score"] == 0 and g["correct"] is False, (blank, g)
+        assert "reference answer" in g["explanation"].lower()
+    assert calls["n"] == 0, "empty answer must not call the model"
+    print("ok  empty answer short-circuits to score 0 (no model call)")
+
+
 if __name__ == "__main__":
     test_clean_json_first_try()
     test_repair_retry_recovers()
     test_safe_default_when_never_valid()
     test_out_of_range_score_rejected()
+    test_empty_answer_short_circuits_to_zero()
     print("\nAll NAH-8 grade_answer tests passed.")