diff --git a/src/helm/benchmark/metrics/codeinsights_code_evaluation_metrics.py b/src/helm/benchmark/metrics/codeinsights_code_evaluation_metrics.py
index acbc80b277..a5cdb1e2d2 100644
--- a/src/helm/benchmark/metrics/codeinsights_code_evaluation_metrics.py
+++ b/src/helm/benchmark/metrics/codeinsights_code_evaluation_metrics.py
@@ -88,7 +88,11 @@ def calculate_ast_distance(self, code1: str, code2: str) -> float:
             nodes1: List[str] = self._extract_ast_features(tu1.cursor)
             nodes2: List[str] = self._extract_ast_features(tu2.cursor)
 
-            return levenshtein_distance_ratio(nodes1, nodes2)
+            # `Levenshtein.ratio` is a similarity in [0, 1] (1.0 = identical),
+            # while the docstring, the public stat name "ast_distance", and the
+            # error-fallback `return 1.0` below all assume *distance* semantics
+            # (0 = identical, 1 = completely different). Convert to a distance.
+            return 1.0 - levenshtein_distance_ratio(nodes1, nodes2)
 
         except Exception:
             # any parse error or clang error → max distance
@@ -232,7 +236,10 @@ def evaluate_generation(
         if gen_asm == "" or truth_asm == "":
             asm_distance = 1.0
         else:
-            asm_distance = levenshtein_distance_ratio(gen_asm, truth_asm)
+            # `Levenshtein.ratio` returns a similarity in [0, 1]; convert to
+            # distance so the value matches the "asm_distance" stat name and
+            # the compile-failure fallback above (1.0 = max distance).
+            asm_distance = 1.0 - levenshtein_distance_ratio(gen_asm, truth_asm)
 
         # Create assembly-based statistics
         stats.extend(self._create_asm_stats(asm_distance))
diff --git a/src/helm/benchmark/metrics/test_codeinsights_code_evaluation_metrics.py b/src/helm/benchmark/metrics/test_codeinsights_code_evaluation_metrics.py
new file mode 100644
index 0000000000..7f3e6299c9
--- /dev/null
+++ b/src/helm/benchmark/metrics/test_codeinsights_code_evaluation_metrics.py
@@ -0,0 +1,64 @@
+"""Tests for codeinsights_code_evaluation_metrics distance semantics.
+
+`Levenshtein.ratio` returns a *similarity* in [0, 1] (1.0 = identical), but the
+metric is reported as ``ast_distance`` / ``asm_distance`` and its error
+fallback returns 1.0 for "max distance". Computing similarity directly meant a
+perfect match was reported as 1.0 — indistinguishable from a parse failure.
+These tests pin the distance semantics: identical inputs return 0.0 and a
+totally-different input returns 1.0.
+"""
+
+import pytest
+
+
+# The module imports clang.cindex and Levenshtein at top level (both ship in
+# the codeinsights extras). Skip cleanly when they're absent.
+clang_cindex = pytest.importorskip("clang.cindex")
+pytest.importorskip("Levenshtein")
+
+
+def _patch_clang_index(monkeypatch):
+    """Replace clang.cindex.Index.create so ASTAnalyzer.__init__ doesn't need
+    libclang installed; ``parse`` round-trips the source text through .cursor
+    so the stubbed _extract_ast_features can read it back."""
+
+    class _DummyTU:
+        def __init__(self, code: str) -> None:
+            self.cursor = code
+
+    class _DummyIndex:
+        def parse(self, path, args, unsaved_files):
+            return _DummyTU(unsaved_files[0][1])
+
+    monkeypatch.setattr(clang_cindex.Index, "create", staticmethod(lambda: _DummyIndex()))
+
+
+def test_calculate_ast_distance_identical_returns_zero(monkeypatch):
+    _patch_clang_index(monkeypatch)
+    from helm.benchmark.metrics.codeinsights_code_evaluation_metrics import ASTAnalyzer
+
+    analyzer = ASTAnalyzer()
+    analyzer._extract_ast_features = lambda cursor: list(cursor)
+
+    assert analyzer.calculate_ast_distance("int x = 1;", "int x = 1;") == 0.0
+
+
+def test_calculate_ast_distance_completely_different_returns_one(monkeypatch):
+    _patch_clang_index(monkeypatch)
+    from helm.benchmark.metrics.codeinsights_code_evaluation_metrics import ASTAnalyzer
+
+    analyzer = ASTAnalyzer()
+    analyzer._extract_ast_features = lambda cursor: list(cursor)
+
+    assert analyzer.calculate_ast_distance("aaaa", "zzzz") == 1.0
+
+
+def test_calculate_ast_distance_partial_overlap_is_small_positive(monkeypatch):
+    _patch_clang_index(monkeypatch)
+    from helm.benchmark.metrics.codeinsights_code_evaluation_metrics import ASTAnalyzer
+
+    analyzer = ASTAnalyzer()
+    analyzer._extract_ast_features = lambda cursor: list(cursor)
+
+    distance = analyzer.calculate_ast_distance("int x = 1;", "int y = 2;")
+    assert 0.0 < distance < 1.0