diff --git a/rulechef/ranking.py b/rulechef/ranking.py
index 75a6e78..8b00ff3 100644
--- a/rulechef/ranking.py
+++ b/rulechef/ranking.py
@@ -185,13 +185,19 @@ def prune_harmful_rules(
     rules: list[Rule],
     report: RankingReport,
     min_marginal_f1: float = 0.0,
+    min_support: int = 1,
 ) -> tuple[list[Rule], list[Rule]]:
-    """Split rules into (kept, dropped) based on marginal contribution.
+    """Split rules into (kept, dropped) based on marginal contribution and support.
 
     A rule is dropped when its marginal F1 is known and below
     min_marginal_f1 — i.e. the ensemble measurably does better without it.
     Rules without ablation data are always kept.
 
+    Dead rules are dropped too: a rule whose validated_support is below
+    min_support never fired on the eval split, so its marginal F1 is exactly
+    0.0 and the threshold above can never catch it (issue #4). Pass
+    min_support=0 to keep zero-support rules.
+
     Returns:
         Tuple of (kept_rules, dropped_rules).
     """
@@ -200,8 +206,10 @@ def prune_harmful_rules(
         for r in report.rankings
         if r.marginal_f1 is not None and r.marginal_f1 < min_marginal_f1
     }
-    kept = [r for r in rules if r.id not in harmful_ids]
-    dropped = [r for r in rules if r.id in harmful_ids]
+    dead_ids = {r.id for r in rules if r.validated_support < min_support}
+    drop = harmful_ids | dead_ids
+    kept = [r for r in rules if r.id not in drop]
+    dropped = [r for r in rules if r.id in drop]
     return kept, dropped
 
 
diff --git a/tests/test_ranking.py b/tests/test_ranking.py
index 795ea69..4be9a31 100644
--- a/tests/test_ranking.py
+++ b/tests/test_ranking.py
@@ -148,3 +148,20 @@ def test_keeps_rules_without_ablation_data(self):
         kept, dropped = prune_harmful_rules(rules, report)
         assert len(kept) == 1
         assert not dropped
+
+    def test_prunes_dead_zero_support_rule(self):
+        rules = [
+            _rule("exchange", "exchange_rate", r"exchange rate"),
+            _rule("dead", "card_arrival", r"zebra unicorn"),  # never fires on _dataset()
+        ]
+        report = rank_rules(rules, _dataset(), _apply_rules_fn(), compute_marginal=True)
+        kept, dropped = prune_harmful_rules(rules, report)
+        assert [r.id for r in dropped] == ["dead"]
+        assert {r.id for r in kept} == {"exchange"}
+
+    def test_min_support_zero_keeps_dead_rule(self):
+        rules = [_rule("dead", "card_arrival", r"zebra unicorn")]
+        report = rank_rules(rules, _dataset(), _apply_rules_fn(), compute_marginal=False)
+        kept, dropped = prune_harmful_rules(rules, report, min_support=0)
+        assert len(kept) == 1
+        assert not dropped