diff --git a/CHANGELOG.md b/CHANGELOG.md
index a3babc82..021b3217 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -161,6 +161,9 @@ All notable changes to SkillOpt are documented here. This project adheres to
   scores and gate actions (thanks @Tanmay9223, #96).
 - Add focused coverage for the validation-gate decision core and edit-budget
   schedulers (thanks @nankingjing, #122, #123).
+- Add focused coverage for the DocVQA ANLS evaluator, including the 0.5
+  threshold boundary, the maximum over gold answers, and the accepted gold
+  answer formats (thanks @Shizoqua).
 
 ### Acknowledgements 🙏
 Thank you to the contributors behind this unreleased work:
diff --git a/tests/test_docvqa_evaluator.py b/tests/test_docvqa_evaluator.py
new file mode 100644
index 00000000..a68519f0
--- /dev/null
+++ b/tests/test_docvqa_evaluator.py
@@ -0,0 +1,106 @@
+"""Unit tests for the DocVQA ANLS evaluator.
+
+ANLS (Biten et al., 2019): for each gold answer, NL is the Levenshtein distance
+divided by the longer string's length; the score is 1 - NL when NL < 0.5 and 0
+otherwise, and the item score is the maximum over gold answers.
+"""
+
+import pytest
+
+from skillopt.envs.docvqa.evaluator import (
+    DEFAULT_ANLS_THRESHOLD,
+    _extract_answer_strings,
+    _levenshtein_distance,
+    _score_single_answer,
+    evaluate,
+    extract_answer,
+)
+
+
+@pytest.mark.parametrize(
+    ("a", "b", "distance"),
+    [("kitten", "sitting", 3), ("flaw", "lawn", 2), ("abc", "abc", 0), ("", "abc", 3), ("abc", "", 3)],
+)
+def test_levenshtein_distance_known_values(a, b, distance):
+    assert _levenshtein_distance(a, b) == distance
+    assert _levenshtein_distance(b, a) == distance
+
+
+def test_threshold_is_the_standard_anls_value():
+    assert DEFAULT_ANLS_THRESHOLD == 0.5
+
+
+@pytest.mark.parametrize(
+    ("predicted", "target", "expected"),
+    [
+        ("kitten", "sitting", 1 - 3 / 7),
+        ("abcd", "abce", 0.75),
+        ("abcd", "ab", 0.0),  # NL is exactly 0.5, which is not below the threshold
+        ("abcde", "abcd", 0.8),
+        ("ab", "cd", 0.0),
+    ],
+)
+def test_score_single_answer_follows_anls(predicted, target, expected):
+    assert _score_single_answer(predicted, target, DEFAULT_ANLS_THRESHOLD) == pytest.approx(expected)
+
+
+def test_score_single_answer_empty_strings():
+    assert _score_single_answer("", "", DEFAULT_ANLS_THRESHOLD) == 1.0
+    assert _score_single_answer(None, "", DEFAULT_ANLS_THRESHOLD) == 1.0
+    assert _score_single_answer("x", "", DEFAULT_ANLS_THRESHOLD) == 0.0
+    assert _score_single_answer("", "x", DEFAULT_ANLS_THRESHOLD) == 0.0
+
+
+def test_case_and_whitespace_are_normalized():
+    assert evaluate("<answer>  Total   REVENUE </answer>", ["total revenue"])["anls"] == 1.0
+
+
+@pytest.mark.parametrize("golds", [["2000", "1999", "abcd"], ["abcd", "1999", "2000"]])
+def test_item_score_is_the_max_over_gold_answers(golds):
+    result = evaluate("<answer>1999</answer>", golds)
+
+    assert result["anls"] == 1.0
+    assert result["predicted_answer"] == "1999"
+    assert result["gold_answers"] == golds
+
+
+def test_near_miss_gets_partial_credit_but_not_exact():
+    # The rollout sets hard = anls >= 0.999 and soft = anls.
+    anls = evaluate("<answer>abce</answer>", ["abcd"])["anls"]
+
+    assert anls == pytest.approx(0.75)
+    assert anls < 0.999
+
+
+@pytest.mark.parametrize(
+    ("raw", "expected"),
+    [
+        (["a", "b"], ["a", "b"]),
+        ('["a", "b"]', ["a", "b"]),
+        ("['a', 'b']", ["a", "b"]),
+        ({"answers": ["x"]}, ["x"]),
+        ({"ground_truth": "y"}, ["y"]),
+        ([{"text": "t"}, {"answer": "u"}, {"value": "v"}], ["t", "u", "v"]),
+        ("plain answer", ["plain answer"]),
+        ("[not json", ["[not json"]),
+        (5, ["5"]),
+        (None, [""]),
+        ("", [""]),
+        ([], [""]),
+    ],
+)
+def test_extract_answer_strings_accepts_dataset_shapes(raw, expected):
+    assert _extract_answer_strings(raw) == expected
+
+
+@pytest.mark.parametrize(
+    ("text", "expected"),
+    [
+        ("reasoning\n<ANSWER>Foo</ANSWER>", "Foo"),
+        ("<answer>x</answer> then <answer>y</answer>", "y"),
+        ("first line\n\nlast line  ", "last line"),
+        ("", ""),
+    ],
+)
+def test_extract_answer(text, expected):
+    assert extract_answer(text) == expected