diff --git a/src/trinity/orchestration/reward.py b/src/trinity/orchestration/reward.py
index 5621e74..80a941b 100644
--- a/src/trinity/orchestration/reward.py
+++ b/src/trinity/orchestration/reward.py
@@ -314,6 +314,11 @@ def normalize_math_answer(ans: str | None) -> str:
     s = re.sub(r"\\d?frac\s*(\d)\s*(\d)", r"\1/\2", s)
     s = s.replace(r"\cdot", "*").replace(r"\times", "*")
     s = re.sub(r"\s+", "", s)
+    # Scientific notation a*10^{b} (after \times/\cdot became *) -> canonical
+    # "aeb" so the numeric path can parse it: "5*10^{3}" -> "5e3" -> 5000.0.
+    # Without this, float()/Fraction both fail and the sympy fallback reads "^"
+    # as XOR, so a correct order-of-magnitude answer scores 0.
+    s = re.sub(r"(\d+(?:\.\d+)?)\*10\^\{?(-?\d+)\}?", r"\1e\2", s)
     # Drop thousands-separator commas so "1,234" compares equal to "1234" (and
     # parses as a number). extract_last_number already strips these, so without
     # this the extract path and the compare path disagree and a correct answer
@@ -321,6 +326,11 @@ def normalize_math_answer(ans: str | None) -> str:
     # leaving set/tuple/interval answers like "(1,2)" untouched.
     s = re.sub(r"(?<=\d),(?=\d{3}(?:\D|$))", "", s)
     s = s.lower()
+    # Drop a leading single-variable assignment ("x=5" -> "5"): grade the value,
+    # not the restated variable. Guarded so "==" (equality) and inequalities like
+    # "x<=5"/"x>=5" stay intact — only a bare "<ident>=" prefix not followed by a
+    # second "=" is cut ("<"/">" are not word chars, so they never precede the "=").
+    s = re.sub(r"^[a-z]\w*=(?!=)", "", s)
     # Canonicalize a pure integer ratio a/b.
     m = re.fullmatch(r"\(?(-?\d+)\)?/\(?(-?\d+)\)?", s)
     if m:
@@ -409,7 +419,9 @@ def _sympy_equal(a: str, b: str) -> bool:
 def _check_math(candidate: str, reference: object) -> bool:
     """True iff the candidate's extracted answer equals the reference."""
     extracted = extract_boxed(candidate)
-    if extracted is None:
+    if extracted is None or not extracted.strip():
+        # No box, or an empty/whitespace \boxed{} that carries no answer: fall
+        # back to the last number in the text, exactly like a missing box.
         extracted = extract_last_number(candidate)
     if extracted is None:
         # Last resort: compare the whole (normalized) candidate.
diff --git a/tests/test_reward_math_forms.py b/tests/test_reward_math_forms.py
new file mode 100644
index 0000000..2ef8e46
--- /dev/null
+++ b/tests/test_reward_math_forms.py
@@ -0,0 +1,61 @@
+"""Offline unit tests for extra MATH-500 answer forms in the grader (reward.py).
+
+Covers issue #69's sibling class of false negatives (same family as the
+already-fixed thousands-comma bug #35): a correct answer written as a variable
+assignment (``x=5``), in scientific notation (``5\\times10^{3}``), or after an
+empty ``\\boxed{}`` must grade correct — while wrong values, inequalities, and
+empty boxes with no number must still score 0. Pure stdlib (no torch / GPU /
+network), matching the existing ``score_text`` test precedent.
+"""
+from __future__ import annotations
+
+import pytest
+
+from trinity.orchestration import reward as R
+
+
+@pytest.mark.parametrize(
+    "candidate, reference",
+    [
+        # Variable assignment: the value is graded, not the restated variable.
+        (r"\boxed{x=5}", "5"),
+        (r"\boxed{y = 42}", "42"),
+        # Scientific notation (\times / \cdot -> canonical exponent).
+        (r"\boxed{5\times10^{3}}", "5000"),
+        (r"\boxed{1.5\times10^{3}}", "1500"),
+        (r"\boxed{5\cdot10^{3}}", "5000"),
+        (r"\boxed{5\times10^{-3}}", "0.005"),
+        # Empty \boxed{}: fall back to the last number in the text.
+        (r"so the answer is \boxed{}. It is 42", "42"),
+        (r"\boxed{} 1/2", "0.5"),
+    ],
+)
+def test_extra_answer_forms_grade_correct(candidate, reference):
+    assert R.score_text("math500", candidate, reference) == 1.0
+
+
+@pytest.mark.parametrize(
+    "candidate, reference",
+    [
+        (r"\boxed{x=6}", "5"),      # wrong value, assignment stripped
+        (r"\boxed{x<=5}", "5"),     # inequality must NOT be graded as its RHS
+        (r"\boxed{x>=5}", "5"),
+        (r"\boxed{}", "42"),        # empty box, no number anywhere -> wrong
+        (r"\boxed{6\times10^{3}}", "5000"),  # wrong magnitude
+    ],
+)
+def test_extra_answer_forms_reject_wrong(candidate, reference):
+    assert R.score_text("math500", candidate, reference) == 0.0
+
+
+def test_normalize_drops_only_leading_assignment():
+    assert R.normalize_math_answer("x=5") == "5"
+    # Equality assertion and inequalities are left intact (not turned into "=5").
+    assert R.normalize_math_answer("x==5") != "5"
+    assert R.normalize_math_answer("x<=5") != "5"
+
+
+def test_normalize_rewrites_scientific_notation():
+    # \times becomes * upstream; the exponent form parses numerically.
+    assert R._as_number(R.normalize_math_answer(r"5\times10^{3}")) == 5000.0
+    assert R._as_number(R.normalize_math_answer(r"5\times10^{-3}")) == 0.005