diff --git a/METRICS.md b/METRICS.md
index 02bc62f..26d2b9b 100644
--- a/METRICS.md
+++ b/METRICS.md
@@ -154,6 +154,22 @@ only 10 cases where the prompt differs, a real effect would need to show up as
 several (roughly 3+) net case wins concentrated in those cases. This is a
 judgement from the observed noise, not a significance test; none was computed.
 
+### Confidence signal
+
+Pooling both v2 conditions (144 trials: 131 right #1 picks, 13 wrong):
+
+| Signal | AUC right-vs-wrong (0.5 = coin flip) |
+|---|---|
+| Top-1 `confidence_score` | 0.60 |
+| Gap between #1 and #2 confidence | **0.86** |
+
+Raw confidence is near-useless: wrong picks averaged 0.83–0.89. The Slack card
+and dashboard therefore flag a **close call** when the top two suspects are
+within 0.1, and show both. On these trials that rule flags 9/13 wrong picks
+and 9/131 right ones. The 0.1 cut was chosen on the same 144 trials (in-sample,
+13 wrong picks), so treat it as a working rule. Two wrong picks had a 0.7 gap:
+the signal does not catch confidently wrong rankings.
+
 ### Retrieval
 
 - Runbook top-1 on cases with an expected runbook: **10/17**; correct runbook
diff --git a/core/services/notifier.py b/core/services/notifier.py
index f2ffe44..f6e4917 100644
--- a/core/services/notifier.py
+++ b/core/services/notifier.py
@@ -4,6 +4,16 @@
 
 load_dotenv()
 SLACK_WEBHOOK_URL = os.environ.get("SLACK_WEBHOOK_URL")
+# Raw confidence barely separates right from wrong #1 picks (AUC 0.60 on 144 eval trials); the gap to #2
+# does (AUC 0.86). ponytail: 0.1 was picked on those same trials (13 wrong picks), re-check as data grows.
+CLOSE_CALL_MARGIN = 0.1
+
+
+def is_close_call(commits: list[dict]) -> bool:
+    """Top two suspects within CLOSE_CALL_MARGIN: the #1 pick is not trustworthy on its own."""
+    if len(commits) < 2:
+        return False
+    return round(commits[0]["confidence_score"] - commits[1]["confidence_score"], 2) <= CLOSE_CALL_MARGIN
 
 
 def _fmt_commit(c: dict) -> str:
@@ -49,7 +59,12 @@ def build_incident_card(incident: dict) -> dict:
                 },
             }
         )
-    if commits:
+    if commits and is_close_call(commits):
+        text = ":scales: *Close call: check both suspects.*\n" + "\n".join(
+            f"{i}. " + _fmt_commit(c) for i, c in enumerate(commits[:2], 1)
+        )
+        blocks.append({"type": "section", "text": {"type": "mrkdwn", "text": text}})
+    elif commits:
         blocks.append(
             {
                 "type": "section",
diff --git a/core/tests/test_phase3.py b/core/tests/test_phase3.py
index e77ce45..0052ddf 100644
--- a/core/tests/test_phase3.py
+++ b/core/tests/test_phase3.py
@@ -32,6 +32,23 @@ def test_build_incident_card_includes_error_and_top_suspect():
     assert "Database Connection Pool Exhaustion" in text
 
 
+def _with_suspects(*scores):
+    commits = [
+        {"commit_hash": f"c{i}", "confidence_score": s, "rationale": f"r{i}"} for i, s in enumerate(scores)
+    ]
+    return {**_INCIDENT, "diagnostics": {**_INCIDENT["diagnostics"], "suspect_commits": commits}}
+
+
+def test_close_call_card_shows_both_suspects():
+    text = str(build_incident_card(_with_suspects(0.9, 0.8, 0.1)))  # gap 0.1 (float 0.0999..)
+    assert "Close call" in text and "c0" in text and "c1" in text and "c2" not in text
+
+
+def test_clear_lead_card_shows_only_top_suspect():
+    text = str(build_incident_card(_with_suspects(0.9, 0.7)))
+    assert "Close call" not in text and "c0" in text and "c1" not in text
+
+
 def test_post_incident_to_slack_skips_without_webhook_url():
     with patch("core.services.notifier.SLACK_WEBHOOK_URL", None):
         assert post_incident_to_slack(_INCIDENT) is False
diff --git a/dashboard/src/app/incidents/[id]/page.tsx b/dashboard/src/app/incidents/[id]/page.tsx
index 644ebd2..9513367 100644
--- a/dashboard/src/app/incidents/[id]/page.tsx
+++ b/dashboard/src/app/incidents/[id]/page.tsx
@@ -54,6 +54,11 @@ export default function IncidentDetail() {
         <div className="panel">
           <h2>Suspect commits</h2>
           {diagnostics.suspect_commits.length === 0 && <p>None found in the alert window.</p>}
+          {/* mirrors core/services/notifier.is_close_call: the gap to #2 predicts wrong picks, raw confidence doesn't */}
+          {diagnostics.suspect_commits.length >= 2 &&
+            Math.round((diagnostics.suspect_commits[0].confidence_score - diagnostics.suspect_commits[1].confidence_score) * 100) <= 10 && (
+              <p className="error-sig">Close call: the top two suspects are within 10 points. Check both.</p>
+            )}
           {diagnostics.suspect_commits.map((c) => (
             <div key={c.commit_hash} style={{ marginBottom: 12 }}>
               <p>