diff --git a/src/commander_ai_lab/sim/engine.py b/src/commander_ai_lab/sim/engine.py
index 7ca911d..0daa3eb 100644
--- a/src/commander_ai_lab/sim/engine.py
+++ b/src/commander_ai_lab/sim/engine.py
@@ -60,6 +60,9 @@ class GameEngine:
   ``weights`` explicitly to override (e.g. for A/B testing).
   """
 
+  # London mulligan: max times a player may mulligan before forced keep
+  MAX_MULLIGANS = 4
+
   def __init__(
     self,
     max_turns: int = 25,
@@ -67,6 +70,7 @@ class GameEngine:
     weights: Optional[dict] = None,
     record_log: bool = False,
     ml_log: bool = False,
+    mulligan_rule: str = "london",
   ):
     self.max_turns = max_turns
     self.starting_life = starting_life
@@ -75,6 +79,7 @@ class GameEngine:
     self.record_log = record_log
     self.ml_log = ml_log
     self.ml_decisions: list[dict] = []
+    self.mulligan_rule = mulligan_rule
 
   # ──────────────────────────────────────────────────────────
   # Public API
@@ -282,6 +287,42 @@ class GameEngine:
   # Internal helpers
   # ──────────────────────────────────────────────────────────
 
+  # ──────────────────────────────────────────────────────────
+  # Opening-hand helpers
+  # ──────────────────────────────────────────────────────────
+
+  @staticmethod
+  def _count_lands_in_hand(hand: list[Card]) -> int:
+    """Count lands in a hand of cards."""
+    return sum(1 for c in hand if c.is_land())
+
+  @staticmethod
+  def _should_mulligan(hand: list[Card], mulligan_number: int) -> bool:
+    """Heuristic: mulligan if opening hand has 0-1 lands or 6-7 lands.
+
+    Becomes more lenient with each successive mulligan — after 2+
+    mulligans, accept 1-land hands to avoid going too low on cards.
+    """
+    lands = GameEngine._count_lands_in_hand(hand)
+    total = len(hand)
+    if mulligan_number >= 2:
+      # Desperate: keep anything with 1-5 lands
+      return lands < 1 or lands >= total
+    # Standard: keep 2-5 lands in a 7-card hand
+    return lands <= 1 or lands >= 6
+
+  @staticmethod
+  def _pick_bottom_cards(hand: list[Card], count: int, weights: dict) -> list[Card]:
+    """Choose the worst *count* cards to put on the bottom of library.
+
+    Uses the AI scoring function so the least valuable cards are
+    bottomed — mirrors a real player's London mulligan decision.
+    """
+    if count <= 0 or count >= len(hand):
+      return []
+    scored = sorted(hand, key=lambda c: score_card(c, weights))
+    return scored[:count]
+
   def _create_state(
     self,
     decks: list[list[Card]],
@@ -290,6 +331,11 @@ class GameEngine:
   ) -> SimState:
     """Create initial simulation state with shuffled decks and opening hands (N players).
 
+    Implements the London mulligan rule: each player draws 7, decides
+    whether to keep, and if not shuffles back and redraws.  After
+    keeping, they put N cards on the bottom of their library where N
+    is the number of times they mulliganed.
+
     If commander_names is provided, the matching card is flagged
     is_commander=True and placed in the player's command_zone
     instead of the library/hand.
@@ -308,16 +354,46 @@ class GameEngine:
           commander_card = c
         else:
           remaining.append(c)
-      random.shuffle(remaining)
-      hand = remaining[:7]
+
+      # ── London mulligan loop ──
+      mulligan_count = 0
+      if self.mulligan_rule == "london":
+        for attempt in range(self.MAX_MULLIGANS):
+          random.shuffle(remaining)
+          hand = remaining[:7]
+          if not self._should_mulligan(hand, attempt):
+            break
+          mulligan_count += 1
+        else:
+          # Exhausted all mulligans — forced keep with last draw
+          random.shuffle(remaining)
+          hand = remaining[:7]
+      else:
+        # No mulligan (legacy behavior)
+        random.shuffle(remaining)
+        hand = remaining[:7]
+
       library = remaining[7:]
+
+      # ── Bottom N cards for London mulligan ──
+      if mulligan_count > 0:
+        bottom = self._pick_bottom_cards(hand, mulligan_count, self.weights)
+        for card in bottom:
+          hand.remove(card)
+          library.insert(0, card)  # bottom of library
+
+      opening_lands = self._count_lands_in_hand(hand)
       player = Player(
         name=name,
         life=self.starting_life,
         owner_id=idx,
         library=library,
         hand=hand,
-        stats=PlayerStats(cards_drawn=7),
+        stats=PlayerStats(
+          cards_drawn=7,
+          mulligans=mulligan_count,
+          opening_hand_lands=opening_lands,
+        ),
       )
       if commander_card:
         player.command_zone.append(commander_card)
diff --git a/src/commander_ai_lab/sim/models.py b/src/commander_ai_lab/sim/models.py
index bf24f1f..adaf299 100644
--- a/src/commander_ai_lab/sim/models.py
+++ b/src/commander_ai_lab/sim/models.py
@@ -102,6 +102,8 @@ class PlayerStats:
     """Per-game statistics tracked for a player."""
 
     cards_drawn: int = 7  # starting hand
+    mulligans: int = 0
+    opening_hand_lands: int = 0
     lands_played: int = 0
     spells_cast: int = 0
     creatures_played: int = 0
@@ -117,6 +119,8 @@ class PlayerStats:
     def to_dict(self) -> dict:
         return {
             "cardsDrawn": self.cards_drawn,
+            "mulligans": self.mulligans,
+            "openingHandLands": self.opening_hand_lands,
             "landsPlayed": self.lands_played,
             "spellsCast": self.spells_cast,
             "creaturesPlayed": self.creatures_played,
diff --git a/tests/sim/test_mulligan.py b/tests/sim/test_mulligan.py
new file mode 100644
index 0000000..768393e
--- /dev/null
+++ b/tests/sim/test_mulligan.py
@@ -0,0 +1,315 @@
+"""Tests for London mulligan and opening-hand validation (#86)."""
+import random
+
+import pytest
+
+from commander_ai_lab.sim.engine import GameEngine
+from commander_ai_lab.sim.models import Card, PlayerStats
+from commander_ai_lab.sim.rules import enrich_card
+
+
+# ── Helpers ───────────────────────────────────────────────────
+
+def _land(name: str = "Forest") -> Card:
+    return enrich_card(Card(name=name))
+
+
+def _spell(name: str = "Grizzly Bears") -> Card:
+    return enrich_card(Card(name=name))
+
+
+def _make_deck(lands: int = 24, spells: int = 36) -> list[Card]:
+    """Build a deck with an exact land/spell split."""
+    deck: list[Card] = []
+    for _ in range(lands):
+        deck.append(_land())
+    for _ in range(spells):
+        deck.append(_spell())
+    return deck
+
+
+def _all_lands_deck(size: int = 60) -> list[Card]:
+    """Deck of nothing but lands — forces 7-land opening hand."""
+    return [_land() for _ in range(size)]
+
+
+def _no_lands_deck(size: int = 60) -> list[Card]:
+    """Deck with zero lands — forces 0-land opening hand."""
+    return [_spell(f"Spell_{i}") for i in range(size)]
+
+
+# ── PlayerStats fields ────────────────────────────────────────
+
+class TestPlayerStatsMultiganFields:
+    def test_default_mulligans_zero(self):
+        stats = PlayerStats()
+        assert stats.mulligans == 0
+
+    def test_default_opening_hand_lands_zero(self):
+        stats = PlayerStats()
+        assert stats.opening_hand_lands == 0
+
+    def test_to_dict_includes_mulligan_keys(self):
+        stats = PlayerStats(mulligans=2, opening_hand_lands=3)
+        d = stats.to_dict()
+        assert d["mulligans"] == 2
+        assert d["openingHandLands"] == 3
+
+
+# ── _should_mulligan heuristic ────────────────────────────────
+
+class TestShouldMulligan:
+    def test_zero_lands_mulligans(self):
+        hand = [_spell() for _ in range(7)]
+        assert GameEngine._should_mulligan(hand, 0) is True
+
+    def test_one_land_mulligans(self):
+        hand = [_land()] + [_spell() for _ in range(6)]
+        assert GameEngine._should_mulligan(hand, 0) is True
+
+    def test_two_lands_keeps(self):
+        hand = [_land(), _land()] + [_spell() for _ in range(5)]
+        assert GameEngine._should_mulligan(hand, 0) is False
+
+    def test_five_lands_keeps(self):
+        hand = [_land() for _ in range(5)] + [_spell(), _spell()]
+        assert GameEngine._should_mulligan(hand, 0) is False
+
+    def test_six_lands_mulligans(self):
+        hand = [_land() for _ in range(6)] + [_spell()]
+        assert GameEngine._should_mulligan(hand, 0) is True
+
+    def test_seven_lands_mulligans(self):
+        hand = [_land() for _ in range(7)]
+        assert GameEngine._should_mulligan(hand, 0) is True
+
+    def test_desperate_keeps_one_land(self):
+        """After 2+ mulligans, accept a 1-land hand."""
+        hand = [_land()] + [_spell() for _ in range(6)]
+        assert GameEngine._should_mulligan(hand, 2) is False
+
+
+# ── _count_lands_in_hand ──────────────────────────────────────
+
+class TestCountLandsInHand:
+    def test_empty_hand(self):
+        assert GameEngine._count_lands_in_hand([]) == 0
+
+    def test_mixed_hand(self):
+        hand = [_land(), _spell(), _land(), _spell()]
+        assert GameEngine._count_lands_in_hand(hand) == 2
+
+    def test_all_lands(self):
+        hand = [_land() for _ in range(7)]
+        assert GameEngine._count_lands_in_hand(hand) == 7
+
+
+# ── _pick_bottom_cards ────────────────────────────────────────
+
+class TestPickBottomCards:
+    def test_returns_correct_count(self):
+        hand = [_land(), _land(), _spell(), _spell(), _spell()]
+        bottom = GameEngine._pick_bottom_cards(hand, 2, {})
+        assert len(bottom) == 2
+
+    def test_zero_count_returns_empty(self):
+        hand = [_land(), _spell()]
+        assert GameEngine._pick_bottom_cards(hand, 0, {}) == []
+
+    def test_count_gte_hand_returns_empty(self):
+        hand = [_land()]
+        assert GameEngine._pick_bottom_cards(hand, 1, {}) == []
+
+    def test_bottoms_lowest_scored_cards(self):
+        """The card with the worst AI score should be picked to bottom."""
+        good = enrich_card(Card(name="Serra Angel", type_line="Creature", cmc=5, pt="4/4"))
+        bad = enrich_card(Card(name="Forest"))
+        hand = [good, bad, _spell()]
+        bottom = GameEngine._pick_bottom_cards(hand, 1, {})
+        # Lands score very low compared to creatures
+        assert bottom[0].name == "Forest"
+
+
+# ── London mulligan integration ───────────────────────────────
+
+class TestLondonMulliganIntegration:
+    def test_good_hand_no_mulligan(self):
+        """A well-balanced deck should usually keep on the first try."""
+        random.seed(42)
+        deck = _make_deck(lands=24, spells=36)
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run(
+            deck_a=[c.clone() for c in deck],
+            deck_b=[c.clone() for c in deck],
+        )
+        # At least one player should have kept without mulligan
+        any_kept = any(
+            p.stats.mulligans == 0 for p in result.players
+        )
+        # With a 24-land deck, most seeds give keepable hands
+        assert result.turns > 0  # game still works
+
+    def test_all_lands_deck_triggers_mulligan(self):
+        """A 60-land deck should always mulligan (7 lands in hand)."""
+        random.seed(99)
+        deck_lands = _all_lands_deck()
+        deck_normal = _make_deck()
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run(
+            deck_a=[c.clone() for c in deck_lands],
+            deck_b=[c.clone() for c in deck_normal],
+        )
+        # Player A has an all-land deck; must have mulliganed at least once
+        pa_stats = result.players[0].stats
+        assert pa_stats.mulligans > 0
+
+    def test_no_lands_deck_triggers_mulligan(self):
+        """A 0-land deck should always mulligan (0 lands in hand)."""
+        random.seed(77)
+        deck_empty = _no_lands_deck()
+        deck_normal = _make_deck()
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run(
+            deck_a=[c.clone() for c in deck_empty],
+            deck_b=[c.clone() for c in deck_normal],
+        )
+        pa_stats = result.players[0].stats
+        assert pa_stats.mulligans > 0
+
+    def test_mulligan_caps_at_max(self):
+        """Even a terrible deck can mulligan at most MAX_MULLIGANS times."""
+        random.seed(11)
+        deck_empty = _no_lands_deck()
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run(
+            deck_a=[c.clone() for c in deck_empty],
+            deck_b=[c.clone() for c in deck_empty],
+        )
+        for p in result.players:
+            assert p.stats.mulligans <= GameEngine.MAX_MULLIGANS
+
+    def test_hand_size_after_mulligan(self):
+        """After N mulligans, opening hand should be 7 - N cards."""
+        random.seed(55)
+        deck_lands = _all_lands_deck()
+        deck_normal = _make_deck()
+        engine = GameEngine(mulligan_rule="london")
+        # We need to test the internal state, so use _create_state directly
+        sim = engine._create_state(
+            decks=[[c.clone() for c in deck_lands], [c.clone() for c in deck_normal]],
+            names=["LandPlayer", "NormalPlayer"],
+        )
+        for p in sim.players:
+            expected = 7 - p.stats.mulligans
+            assert len(p.hand) == expected, (
+                f"{p.name}: hand={len(p.hand)}, mulligans={p.stats.mulligans}, "
+                f"expected={expected}"
+            )
+
+    def test_opening_hand_lands_tracked(self):
+        """opening_hand_lands should reflect final kept hand."""
+        random.seed(42)
+        deck = _make_deck()
+        engine = GameEngine(mulligan_rule="london")
+        sim = engine._create_state(
+            decks=[[c.clone() for c in deck], [c.clone() for c in deck]],
+            names=["P1", "P2"],
+        )
+        for p in sim.players:
+            actual_lands = sum(1 for c in p.hand if c.is_land())
+            assert p.stats.opening_hand_lands == actual_lands
+
+    def test_total_cards_preserved(self):
+        """Mulligan must not create or destroy cards."""
+        random.seed(33)
+        deck = _make_deck(lands=24, spells=36)
+        engine = GameEngine(mulligan_rule="london")
+        for _ in range(10):
+            sim = engine._create_state(
+                decks=[[c.clone() for c in deck], [c.clone() for c in deck]],
+                names=["A", "B"],
+            )
+            for p in sim.players:
+                total = len(p.hand) + len(p.library) + len(p.command_zone)
+                assert total == 60, f"Card count mismatch: {total}"
+
+
+# ── mulligan_rule="none" (legacy) ─────────────────────────────
+
+class TestNoMulliganMode:
+    def test_no_mulligan_always_keeps_seven(self):
+        """With mulligan_rule='none', always get 7 cards, 0 mulligans."""
+        random.seed(42)
+        deck = _all_lands_deck()
+        engine = GameEngine(mulligan_rule="none")
+        sim = engine._create_state(
+            decks=[[c.clone() for c in deck]],
+            names=["P1"],
+        )
+        p = sim.players[0]
+        assert len(p.hand) == 7
+        assert p.stats.mulligans == 0
+
+
+# ── N-player mulligan ─────────────────────────────────────────
+
+class TestNPlayerMulligan:
+    def test_four_player_mulligan(self):
+        """Each player in a 4-player game resolves mulligans independently."""
+        random.seed(42)
+        decks = [_make_deck() for _ in range(4)]
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run_n(
+            decks=[[c.clone() for c in d] for d in decks],
+            names=["A", "B", "C", "D"],
+        )
+        assert len(result.players) == 4
+        for p in result.players:
+            assert p.stats.mulligans >= 0
+            assert p.stats.mulligans <= GameEngine.MAX_MULLIGANS
+
+
+# ── Commander + mulligan interaction ──────────────────────────
+
+class TestCommanderMulligan:
+    def test_commander_stays_in_command_zone_through_mulligans(self):
+        """Commander should remain in command zone regardless of mulligans."""
+        random.seed(42)
+        # Build a deck with a known commander
+        deck = _make_deck(lands=24, spells=35)
+        commander = Card(name="Kenrith, the Returned King", type_line="Creature", cmc=5, pt="5/5")
+        deck.append(commander)
+        engine = GameEngine(mulligan_rule="london")
+        sim = engine._create_state(
+            decks=[[c.clone() for c in deck]],
+            names=["P1"],
+            commander_names=["Kenrith, the Returned King"],
+        )
+        p = sim.players[0]
+        assert len(p.command_zone) == 1
+        assert p.command_zone[0].name == "Kenrith, the Returned King"
+        assert p.command_zone[0].is_commander is True
+        # Commander should not appear in hand or library
+        assert all(not c.is_commander for c in p.hand)
+        assert all(not c.is_commander for c in p.library)
+
+
+# ── Serialization round-trip ──────────────────────────────────
+
+class TestMulliganSerialization:
+    def test_result_dict_includes_mulligans(self):
+        """to_dict() output should include mulligan stats."""
+        import json
+        random.seed(42)
+        deck = _make_deck()
+        engine = GameEngine(mulligan_rule="london")
+        result = engine.run(
+            deck_a=[c.clone() for c in deck],
+            deck_b=[c.clone() for c in deck],
+        )
+        d = result.to_dict()
+        # Verify JSON-serializable
+        json.dumps(d)
+        for pr in d["playerResults"]:
+            assert "mulligans" in pr["stats"]
+            assert "openingHandLands" in pr["stats"]