diff --git a/src/interscript/engine.py b/src/interscript/engine.py
index 0d4afad..a7172a1 100644
--- a/src/interscript/engine.py
+++ b/src/interscript/engine.py
@@ -17,7 +17,7 @@
 import re
 import unicodedata
 
-from .expr import expr_to_literal, expr_to_regex, is_plain_string
+from .expr import expr_max_length, expr_to_literal, expr_to_regex, is_plain_string
 
 
 class ExecutionError(ValueError):
@@ -38,13 +38,18 @@ def _compile_parallel(subs: list[dict]) -> tuple[re.Pattern[str], dict[str, str]
             full = "(?<=" + expr_to_regex(sub["before"]) + ")" + full
         if sub.get("after"):
             full = full + "(?=" + expr_to_regex(sub["after"]) + ")"
-        indexed.append((pat, full, f"s{i}"))
+        key = expr_max_length(sub["pattern"])
+        for guard in ("before", "after", "not_before", "not_after"):
+            if sub.get(guard):
+                key += expr_max_length(sub[guard])
+        key += sub.get("priority", 0)
+        indexed.append((key, full, f"s{i}"))
         if is_plain_string(sub["pattern"]) and not sub.get("before") and not sub.get("after"):
             src = expr_to_literal(sub["pattern"])
             if src.upper() != src:
                 anchor_results[f"a{i}"] = expr_to_literal(sub["result"])
-                indexed.append((re.escape(src.upper()), re.escape(src.upper()), f"a{i}"))
-    indexed.sort(key=lambda t: -len(t[0]))
+                indexed.append((len(src), re.escape(src.upper()), f"a{i}"))
+    indexed.sort(key=lambda t: -t[0])
     combined = "|".join(f"(?P<{name}>{full})" for _, full, name in indexed)
     pattern = re.compile(combined) if indexed else re.compile(r"(?!)")
 
diff --git a/src/interscript/expr.py b/src/interscript/expr.py
index 7b21151..ddbc139 100644
--- a/src/interscript/expr.py
+++ b/src/interscript/expr.py
@@ -123,3 +123,23 @@ def expr_to_literal(expr: str) -> str:
 
 def is_plain_string(expr: str) -> bool:
     return bool(re.fullmatch(r'"(?:[^"\\]|\\.)*"', expr.strip()))
+
+
+def expr_max_length(expr: str) -> int:
+    """The Ruby runtime's parallel-selection key: Rule::Sub#max_length =
+    from + before + after + not_before + not_after (+ priority), where a
+    zero-width stdlib alias (boundary, line_start, ...) counts 1."""
+    total = 0
+    for kind, value in _scan(expr, "expression"):
+        if kind == "lit":
+            total += len(value)
+        elif kind in ("cls", "range", "space", "boundary", "nwb", "anchor"):
+            total += 1
+        elif kind == "alt":
+            total += max(len(a) for a in value.split("\x00"))
+        elif kind == "opt":
+            total += len(value)
+        elif kind == "grp":
+            total += expr_max_length(value)
+        # cat: concatenation marker
+    return total
diff --git a/tests/test_engine.py b/tests/test_engine.py
index ecdb104..2aa1c11 100644
--- a/tests/test_engine.py
+++ b/tests/test_engine.py
@@ -115,3 +115,23 @@ def test_engine_separate_and_title_case():
     )
     assert transliterate("tc", "hello world") == "Hello world"
     assert transliterate("tc", "hello world\nhello hello") == "Hello world\nHello hello"
+
+
+def test_parallel_selection_matches_ruby_max_length():
+    """Selection mirrors Ruby Rule::Sub#max_length: from + all guard
+    lengths (+priority), zero-width aliases counting 1. A longer later
+    rule (4) must beat an earlier boundary-guarded one (3) — the
+    alalc-ara hamza shape ("\u0623\u064e" vs boundary+"\u0623")."""
+    tree = parse_imp(
+        'stage {\n  parallel {\n'
+        '    sub boundary + "ab", "X"\n'
+        '    sub "abcd", "Y"\n  }\n}\n'
+    )
+    assert Engine(tree).transliterate("abcd") == "Y"
+    # Equal keys keep source order, like Ruby's index tiebreak.
+    tree2 = parse_imp(
+        'stage {\n  parallel {\n'
+        '    sub "ab", "X"\n'
+        '    sub "abc", "Y"\n  }\n}\n'
+    )
+    assert Engine(tree2).transliterate("abc") == "Y"