diff --git a/docs/adr/0001-episode-semantics-boundaries.md b/docs/adr/0001-episode-semantics-boundaries.md
index 44074e1..6564c6e 100644
--- a/docs/adr/0001-episode-semantics-boundaries.md
+++ b/docs/adr/0001-episode-semantics-boundaries.md
@@ -151,10 +151,20 @@ does not change the ordered oracle. This accepts the candidate as compatible
 with already-correct isolated segmentation for this bounded null control; it
 does not expand the candidate's production scope.
 
-Together, the two fixtures accept the recovery and null-control hypotheses for
-their bounded cases only. Maximal-window ties, shared evidence, background
-calibration, and a production complexity design still require independent
-evidence. `Detector::analyze()` and `loglens.report.v3` remain unchanged.
+A focused maximal-window tie control uses six events at offsets 0, 1, 2, 3,
+600, and 601 seconds. With threshold five and the inclusive 600-second window,
+the only candidates are `line:1` through `line:5` and `line:2` through
+`line:6`. Both cover five events, span 600 seconds, and require one episode, so
+the chronological window key is the first objective that can distinguish them.
+The earlier candidate wins for both forward and reversed candidate input. This
+accepts the final tie-break and candidate-order invariance for that bounded
+control; it does not settle the separate shared-evidence policy.
+
+Together, the two fixtures and focused tie control accept the recovery,
+null-control, and deterministic tie-break hypotheses for their bounded cases
+only. Shared evidence, background calibration, and a production complexity
+design still require independent evidence. `Detector::analyze()` and
+`loglens.report.v3` remain unchanged.
 
 ## Alternatives considered
 
diff --git a/tests/test_episode_candidate_core.py b/tests/test_episode_candidate_core.py
index 6cd2717..5155d16 100644
--- a/tests/test_episode_candidate_core.py
+++ b/tests/test_episode_candidate_core.py
@@ -69,6 +69,26 @@ def test_overlapping_candidates_materialize_one_maximal_window(self) -> None:
         self.assertEqual(len(selected), 1)
         self.assertEqual(selected[0].event_ids, tuple(f"line:{i}" for i in range(1, 7)))
 
+    def test_equal_score_windows_prefer_chronological_key(self) -> None:
+        candidates = enumerate_candidate_windows(
+            make_events([0, 1, 2, 3, 600, 601]), 5, 600
+        )
+        expected = [tuple(f"line:{i}" for i in range(1, 6))]
+
+        self.assertEqual(
+            [candidate.event_ids for candidate in candidates],
+            [
+                tuple(f"line:{i}" for i in range(1, 6)),
+                tuple(f"line:{i}" for i in range(2, 7)),
+            ],
+        )
+        for ordered_candidates in (candidates, list(reversed(candidates))):
+            selected = select_window_separated_candidates(ordered_candidates, 600)
+
+            self.assertEqual(
+                [candidate.event_ids for candidate in selected], expected
+            )
+
     def test_research_size_limits_fail_closed(self) -> None:
         with self.assertRaisesRegex(ValueError, "at most 200 events"):
             enumerate_candidate_windows(make_events(list(range(201))), 5, 600)