Sync from GitHub via hub-sync

learning_engine.py

@@ -7,6 +7,8 @@ app.py depends on them.
 """
 from __future__ import annotations
 
+import re
+
 import llm
 from schema import (
     Card, GradeResult, Session, new_card, new_card_state, new_grade, validate_card,
@@ -77,28 +79,27 @@ def grade_answer(card: Card, user_answer: str) -> GradeResult:
                 else f"Not quite. Expected something like: {card['answer']}")
         return new_grade(score, expl, missed_concept=card["topic"])
 
+    # Tone instruction is kept short and the strict "ONLY JSON" requirement is
+    # reasserted as the LAST thing the model reads (in the user turn) — a verbose
+    # "write warmly" preamble was nudging this small model into prose that didn't
+    # parse, and recency improves format compliance.
     messages = [
         {"role": "system", "content":
-         "You grade a student's answer against a reference answer. "
-         "Scoring rubric (be strict): 0-1 = wrong or names the wrong thing; "
-         "2 = partially relevant but misses the key idea; 3 = mostly correct with "
-         "a minor gap; 4-5 = correct and complete. If the answer is factually "
-         "wrong, score 0-2 (NOT 3). "
-         "Score strictly, but write the feedback warmly. Speak DIRECTLY to the "
-         "student in the second person ('you', 'your answer') with an "
-         "encouraging coaching tone — acknowledge what they got right before what "
-         "they missed. NEVER write in the third person or refer to them as 'the "
-         "student' / 'the student's response'. "
-         "Return ONLY a JSON object with keys: "
-         "score (integer 0-5), explanation (string spoken to the student), "
-         "missed_concept (short string naming what they got wrong, or \"\"). "
-         "Example (return ONE object exactly like this, no other text):\n"
-         '{"score": 1, "explanation": "Good instinct, but that\'s the wrong '
-         'location — the Calvin cycle actually runs in the stroma. Try linking '
-         'it to where the enzymes sit.", "missed_concept": "the specific location"}'},
+         "You grade a student's answer against a reference answer.\n"
+         "Scoring (be strict): 0-1 = wrong / names the wrong thing; 2 = partially "
+         "relevant but misses the key idea; 3 = mostly correct, minor gap; "
+         "4-5 = correct and complete. A factually wrong answer is 0-2, never 3.\n"
+         "Write the feedback warmly, speaking to the learner as \"you\" (never "
+         "\"the student\"): note what's right, then what's missing.\n"
+         "JSON keys: score (0-5 int), explanation (spoken to \"you\"), "
+         "missed_concept (what was wrong, or \"\").\n"
+         "Example: {\"score\": 1, \"explanation\": \"Good instinct, but that's the "
+         "wrong spot — the Calvin cycle runs in the stroma. Tie it to where the "
+         "enzymes sit.\", \"missed_concept\": \"the specific location\"}"},
         {"role": "user", "content":
          f"Question: {card['question']}\nReference answer: {card['answer']}\n"
-         f"Student answer: {user_answer}\nGrade it."},
+         f"Student answer: {user_answer}\n\n"
+         "Grade it. Reply with ONLY the JSON object — no prose, no markdown fences."},
     ]
     # Parser + one repair retry; safe default if the model never returns JSON.
     # A generous 2048-token budget so even a long reasoning preamble can't push
@@ -113,14 +114,36 @@ def grade_answer(card: Card, user_answer: str) -> GradeResult:
             f"reference: {card['answer']}",
             card["topic"],
         )
+    explanation = _to_second_person(str(data.get("explanation", "")).strip())
     return new_grade(
         int(data["score"]),
-        str(data.get("explanation", "")).strip()
-        or f"Reference answer: {card['answer']}",
-        str(data.get("missed_concept") or card["topic"]).strip(),
+        explanation or f"Reference answer: {card['answer']}",
+        _to_second_person(str(data.get("missed_concept") or card["topic"]).strip()),
     )
 
 
+# This small model still slips into the third person ("The student's answer…")
+# perhaps half the time despite the prompt. These swaps are the grammatically
+# SAFE ones — possessives only — so we never mangle subject-verb agreement (we
+# leave "The student identifies…" alone rather than produce "You identifies…").
+_SECOND_PERSON_SUBS = [
+    (re.compile(r"\bthe student'?s answer\b", re.I), "your answer"),
+    (re.compile(r"\bthe student'?s response\b", re.I), "your answer"),
+    (re.compile(r"\bthe student'?s\b", re.I), "your"),
+]
+
+
+def _to_second_person(text: str) -> str:
+    """Rewrite clinical third-person possessives to warm second person, matching
+    the original capitalization ('The student's answer' -> 'Your answer')."""
+    for pat, repl in _SECOND_PERSON_SUBS:
+        text = pat.sub(
+            lambda m, r=repl: r.capitalize() if m.group(0)[:1].isupper() else r,
+            text,
+        )
+    return text
+
+
 def _valid_grade(data) -> bool:
     """A grade is usable only if it carries a numeric, in-range score."""
     if not isinstance(data, dict) or "score" not in data:

test_grade_answer.py

@@ -88,6 +88,27 @@ def test_out_of_range_score_rejected():
     print("ok  out-of-range score rejected -> safe default")
 
 
+def test_third_person_possessive_rewritten_to_second():
+    # The model slips into "The student's answer/response" ~half the time; the
+    # safe possessive swaps are applied to the returned explanation.
+    llm.chat, _ = _fake_chat([
+        '{"score": 1, "explanation": "The student\'s answer, \'magic\', is wrong.", '
+        '"missed_concept": "the student\'s grasp of the mechanism"}'
+    ])
+    g = le.grade_answer(_card(), "magic")
+    assert g["explanation"] == "Your answer, 'magic', is wrong.", g["explanation"]
+    assert g["missed_concept"] == "your grasp of the mechanism", g["missed_concept"]
+    print("ok  third-person possessive rewritten to second person")
+
+
+def test_second_person_leaves_safe_subject_form_alone():
+    # We only swap possessives — a subject "The student identifies..." is left
+    # untouched rather than mangled into "You identifies...".
+    assert le._to_second_person("The student identifies it.") == "The student identifies it."
+    assert le._to_second_person("Your answer is close.") == "Your answer is close."
+    print("ok  subject-form third person left alone (no grammar mangling)")
+
+
 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").
@@ -105,5 +126,7 @@ if __name__ == "__main__":
     test_repair_retry_recovers()
     test_safe_default_when_never_valid()
     test_out_of_range_score_rejected()
+    test_third_person_possessive_rewritten_to_second()
+    test_second_person_leaves_safe_subject_form_alone()
     test_empty_answer_short_circuits_to_zero()
     print("\nAll NAH-8 grade_answer tests passed.")