diff --git a/tests/test_policy_decide.py b/tests/test_policy_decide.py
new file mode 100644
index 0000000..ca95cf4
--- /dev/null
+++ b/tests/test_policy_decide.py
@@ -0,0 +1,530 @@
+"""
+tests/test_policy_decide.py
+
+Round-trip tests for POST /api/policy/decide introduced in Issue #83 Step 2.
+
+Coverage:
+  - Health endpoint sanity
+  - Fast path: GameSession snapshot with state_vector → vector_source="precomputed"
+  - Correct 6177-dim tensor reaches the mock model
+  - All 8 MacroAction values are valid in the response
+  - Fallback path: no state_vector → vector_source="encoder"
+  - Fallback path: wrong-length state_vector → vector_source="encoder"
+  - GameSession field name aliases accepted without 422
+  - 503 returned when model not loaded
+  - /api/policy/stats tracks hit/fallback counters
+  - Phase string variants (MAIN1, main_1, COMBAT) round-trip cleanly
+
+All tests use a MockPolicyInferenceService — no real checkpoint required.
+"""
+from __future__ import annotations
+
+import math
+import types
+from typing import Any, Dict, List, Optional
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from ml.config.scope import NUM_ACTIONS, IDX_TO_ACTION, STATE_DIMS
+
+# ---------------------------------------------------------------------------
+# Constants mirrored from routes/policy.py
+# ---------------------------------------------------------------------------
+EXPECTED_GLOBAL_DIM = 29
+EXPECTED_STATE_DIM = STATE_DIMS.total_state_dim  # 6177
+VALID_MACRO_ACTIONS = {a.value for a in IDX_TO_ACTION.values()}
+
+
+# ---------------------------------------------------------------------------
+# Mock policy service
+# ---------------------------------------------------------------------------
+
+class MockCardEmbeddingIndex:
+    """Returns zero vectors for any card name — no NPZ file needed."""
+    _loaded = True
+    _zero_vec = np.zeros(STATE_DIMS.card_embedding_dim, dtype=np.float32)
+
+    def get_embedding(self, name: str) -> np.ndarray:
+        return self._zero_vec
+
+    def mean_pool_zone(self, card_names: List[str]) -> np.ndarray:
+        return self._zero_vec
+
+
+class MockStateEncoder:
+    """Real StateEncoder wired to MockCardEmbeddingIndex."""
+
+    def __init__(self):
+        from ml.encoder.state_encoder import StateEncoder
+        self._inner = StateEncoder(MockCardEmbeddingIndex())
+        self.dim = self._inner.dim
+
+    def encode(self, decision: dict, playstyle: str = "midrange") -> np.ndarray:
+        return self._inner.encode(decision, playstyle)
+
+    def _encode_zones(self, players: list) -> np.ndarray:
+        return self._inner._encode_zones(players)
+
+    def _encode_playstyle(self, playstyle: str) -> np.ndarray:
+        return self._inner._encode_playstyle(playstyle)
+
+
+class MockPolicyInferenceService:
+    """
+    Stand-in for PolicyInferenceService.
+
+    Records every tensor shape passed to the mock model so tests can
+    assert on dimensions without loading a real checkpoint.
+
+    Always returns action_index=0 (cast_creature) with confidence=0.9.
+    """
+    _loaded: bool = True
+    device: str = "cpu"
+
+    def __init__(self, loaded: bool = True):
+        self._loaded = loaded
+        self.encoder = MockStateEncoder()
+        self.captured_shapes: List[tuple] = []
+        self._build_mock_model()
+
+    def _build_mock_model(self):
+        """Build a callable mock that records input shapes and returns logits."""
+        captured = self.captured_shapes
+        import torch
+
+        def fake_forward(tensor):
+            captured.append(tuple(tensor.shape))
+            # Return uniform logits favouring action 0
+            logits = torch.zeros(tensor.shape[0], NUM_ACTIONS)
+            logits[0, 0] = 2.0  # cast_creature wins softmax
+            return logits
+
+        self.model = MagicMock(side_effect=fake_forward)
+
+    def predict(self, decision_snapshot: Dict, playstyle: str = "midrange",
+                temperature: float = 1.0, greedy: bool = False) -> Dict:
+        if not self._loaded:
+            return {"error": "Model not loaded", "detail": "mock"}
+        import torch
+        state_vec = self.encoder.encode(decision_snapshot, playstyle)
+        state_tensor = torch.from_numpy(state_vec.astype(np.float32)).unsqueeze(0)
+        self.captured_shapes.append(tuple(state_tensor.shape))
+        # Always return cast_creature
+        return {
+            "action": IDX_TO_ACTION[0].value,
+            "action_index": 0,
+            "confidence": 0.9,
+            "probabilities": {IDX_TO_ACTION[i].value: (0.9 if i == 0 else 0.1 / (NUM_ACTIONS - 1))
+                              for i in range(NUM_ACTIONS)},
+            "inference_ms": 0.1,
+        }
+
+
+# ---------------------------------------------------------------------------
+# App factory
+# ---------------------------------------------------------------------------
+
+def _make_app(loaded: bool = True) -> tuple[FastAPI, MockPolicyInferenceService]:
+    """Create a fresh FastAPI app with the policy router registered."""
+    from routes.policy import register_policy_routes
+    import routes.policy as policy_mod
+
+    # Reset module-level state between tests
+    policy_mod._policy_service = None
+    policy_mod._online_store = None
+    policy_mod._session_stats = {
+        "total_decisions": 0,
+        "total_tuples_collected": 0,
+        "total_rewards_submitted": 0,
+        "games_completed": 0,
+        "session_start": 0.0,
+        "last_decision_ms": 0.0,
+        "precomputed_vector_hits": 0,
+        "encoder_fallbacks": 0,
+    }
+
+    app = FastAPI()
+    svc = MockPolicyInferenceService(loaded=loaded)
+    register_policy_routes(app, svc)
+    return app, svc
+
+
+# ---------------------------------------------------------------------------
+# Fixture: a realistic GameSession snapshot
+# ---------------------------------------------------------------------------
+
+def _game_session_snapshot(
+    state_vector: Optional[List[float]] = None,
+    include_aliases: bool = True,
+) -> Dict[str, Any]:
+    """
+    Mimics GameSession.buildStateSnapshot() output at turn 3, MAIN1.
+    Includes the state_vector field added in Step 1.
+    """
+    sv = state_vector if state_vector is not None else _valid_state_vector()
+    players = [
+        {
+            "seat": 0,
+            "name": "Human",
+            "isAI": False,
+            "deckName": "Atraxa Superfriends",
+            "life": 37,
+            "poison": 0,
+            "commanderTax": 0,
+            "handCount": 6,
+            "hand": ["Doubling Season", "Deepglow Skate", "Sol Ring",
+                     "Atraxa, Praetors' Voice", "Swamp", "Forest"],
+            "battlefield": ["Sol Ring", "Atraxa, Praetors' Voice"],
+            "graveyard": [],
+            "commandZone": ["Atraxa, Praetors' Voice"],
+            "manaPool": {"W": 0, "U": 0, "B": 0, "R": 0, "G": 0, "C": 0},
+        },
+        {
+            "seat": 1,
+            "name": "AI-1",
+            "isAI": True,
+            "deckName": "Ur-Dragon",
+            "life": 40,
+            "poison": 0,
+            "commanderTax": 2,
+            "handCount": 7,
+            "hand": [],
+            "battlefield": ["Dragon's Hoard", "Savage Ventmaw"],
+            "graveyard": ["Swords to Plowshares"],
+            "commandZone": ["The Ur-Dragon"],
+            "manaPool": {"W": 0, "U": 0, "B": 0, "R": 0, "G": 0, "C": 0},
+        },
+    ]
+
+    snap: Dict[str, Any] = {
+        "schema": "1.1.0",
+        "phase": "MAIN1",
+        "turnNumber": 3,
+        "activePlayer": 0,
+        "priorityPlayer": 0,
+        "awaitingInput": True,
+        "players": players,
+        "stack": [],
+        "legalActions": [
+            {"type": "PASS_PRIORITY", "label": "Pass Priority"},
+            {"type": "CAST_SPELL", "cardId": "Sol Ring", "label": "Cast Sol Ring"},
+        ],
+        "playstyle": "midrange",
+        "greedy": False,
+        "state_vector": sv,
+        "state_vector_dim": len(sv),
+    }
+    return snap
+
+
+def _valid_state_vector() -> List[float]:
+    """A plausible float[29] global scalar vector from buildStateVector(seat=0)."""
+    # life=37/40, everything else normalised appropriately
+    sv = [0.0] * 29
+    sv[0] = 37 / 40.0   # self life
+    sv[3] = 0.0          # cmdr_tax
+    sv[4] = 6 / 15.0    # cards_in_hand
+    sv[6] = 2 / 30.0    # creatures
+    sv[7] = 2 / 100.0   # power proxy
+    sv[8] = 1.0          # is_active_player
+    sv[9] = 1.0          # phase_onehot[main1]
+    # opponent block: indices 14..27
+    sv[14] = 40 / 40.0  # opp life
+    sv[17] = 2 / 10.0   # opp cmdr_tax
+    sv[18] = 7 / 15.0   # opp hand
+    sv[20] = 2 / 30.0   # opp creatures
+    # turn
+    sv[28] = 3 / 25.0
+    return sv
+
+
+# ===========================================================================
+# Tests
+# ===========================================================================
+
+class TestPolicyHealth:
+    def test_health_model_loaded(self):
+        app, _ = _make_app(loaded=True)
+        client = TestClient(app)
+        resp = client.get("/api/policy/health")
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["ready"] is True
+        assert data["status"] == "ok"
+        assert data["model_loaded"] is True
+
+    def test_health_model_not_loaded(self):
+        app, _ = _make_app(loaded=False)
+        client = TestClient(app)
+        resp = client.get("/api/policy/health")
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["ready"] is False
+        assert data["status"] == "degraded"
+
+
+class TestDecideFastPath:
+    """Precomputed state_vector present → vector_source=precomputed."""
+
+    def test_precomputed_path_used(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+        data = resp.json()
+        assert data["vector_source"] == "precomputed"
+
+    def test_action_is_valid_macro_action(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["action"] in VALID_MACRO_ACTIONS
+
+    def test_action_index_in_range(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert 0 <= resp.json()["action_index"] < NUM_ACTIONS
+
+    def test_confidence_is_probability(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        conf = resp.json()["confidence"]
+        assert 0.0 <= conf <= 1.0
+
+    def test_log_prob_is_negative(self):
+        """log(p) for p in (0,1] must be <= 0."""
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["log_prob"] <= 0.0
+
+    def test_tensor_shape_is_6177(self):
+        """The mock model must receive a (1, 6177) tensor."""
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        client.post("/api/policy/decide", json=snap)
+        assert len(svc.captured_shapes) > 0
+        shape = svc.captured_shapes[-1]
+        assert shape == (1, EXPECTED_STATE_DIM), (
+            f"Expected (1, {EXPECTED_STATE_DIM}), got {shape}"
+        )
+
+    def test_probabilities_sum_to_one(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        probs = resp.json().get("probabilities", {})
+        if probs:
+            total = sum(probs.values())
+            assert abs(total - 1.0) < 0.01, f"Probs sum to {total}"
+
+    def test_inference_ms_is_positive(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["inference_ms"] >= 0.0
+
+
+class TestDecideFallbackPath:
+    """No / wrong state_vector → falls back to full encoder."""
+
+    def test_fallback_when_no_state_vector(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap.pop("state_vector", None)
+        snap.pop("state_vector_dim", None)
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["vector_source"] == "encoder"
+
+    def test_fallback_when_wrong_length_vector(self):
+        """A 10-float vector is too short — must fall back."""
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot(state_vector=[0.5] * 10)
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["vector_source"] == "encoder"
+
+    def test_fallback_still_returns_valid_action(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap.pop("state_vector", None)
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200
+        assert resp.json()["action"] in VALID_MACRO_ACTIONS
+
+
+class TestGameSessionAliases:
+    """GameSession field name aliases must not cause 422 validation errors."""
+
+    def test_turnNumber_alias(self):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["turnNumber"] = 5
+        snap.pop("turn", None)
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+    def test_activePlayer_alias(self):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["activePlayer"] = 0
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+    def test_commanderTax_in_player(self):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["players"][1]["commanderTax"] = 4
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+    def test_commandZone_alias(self):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["players"][0]["commandZone"] = ["Atraxa, Praetors' Voice"]
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+    def test_legalActions_alias(self):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["legalActions"] = snap.pop("legal_actions", snap.get("legalActions", []))
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+    def test_schema_110_accepted(self):
+        """schema=1.1.0 (GameSession Step 1 bump) is accepted without error."""
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["schema"] = "1.1.0"
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, resp.text
+
+
+class TestPhaseVariants:
+    """Both Forge-style (MAIN1) and ML-style (main_1) phase strings work."""
+
+    @pytest.mark.parametrize("phase", [
+        "MAIN1", "main_1", "MAIN2", "main_2",
+        "BEGIN_COMBAT", "combat", "END", "end", "CLEANUP",
+    ])
+    def test_phase_accepted(self, phase):
+        app, _ = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap["phase"] = phase
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 200, f"Phase '{phase}' failed: {resp.text}"
+
+
+class TestDecideModelNotLoaded:
+    def test_503_when_not_loaded(self):
+        app, _ = _make_app(loaded=False)
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        resp = client.post("/api/policy/decide", json=snap)
+        assert resp.status_code == 503
+        assert "not loaded" in resp.json()["detail"].lower()
+
+
+class TestStats:
+    def test_precomputed_hits_incremented(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        # Fire 3 precomputed requests
+        for _ in range(3):
+            client.post("/api/policy/decide", json=snap)
+        stats = client.get("/api/policy/stats").json()
+        assert stats["precomputed_vector_hits"] == 3
+        assert stats["total_decisions"] == 3
+
+    def test_encoder_fallbacks_incremented(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        snap.pop("state_vector", None)
+        # Fire 2 fallback requests
+        for _ in range(2):
+            client.post("/api/policy/decide", json=snap)
+        stats = client.get("/api/policy/stats").json()
+        assert stats["encoder_fallbacks"] == 2
+        assert stats["precomputed_vector_hits"] == 0
+
+    def test_mixed_hits_and_fallbacks(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap_pre = _game_session_snapshot()                         # has state_vector
+        snap_enc = _game_session_snapshot()
+        snap_enc.pop("state_vector", None)                         # no state_vector
+        client.post("/api/policy/decide", json=snap_pre)
+        client.post("/api/policy/decide", json=snap_enc)
+        client.post("/api/policy/decide", json=snap_pre)
+        stats = client.get("/api/policy/stats").json()
+        assert stats["precomputed_vector_hits"] == 2
+        assert stats["encoder_fallbacks"] == 1
+        assert stats["total_decisions"] == 3
+
+    def test_decisions_per_second_non_negative(self):
+        app, svc = _make_app()
+        client = TestClient(app)
+        snap = _game_session_snapshot()
+        client.post("/api/policy/decide", json=snap)
+        stats = client.get("/api/policy/stats").json()
+        assert stats["decisions_per_second"] >= 0.0
+
+
+class TestStateVectorDimension:
+    """Explicit dim assertions on the precomputed encoding path."""
+
+    def test_global_block_length_is_29(self):
+        sv = _valid_state_vector()
+        assert len(sv) == EXPECTED_GLOBAL_DIM
+
+    def test_life_normalisation(self):
+        sv = _valid_state_vector()
+        assert 0.0 <= sv[0] <= 1.0, "life not normalised"
+
+    def test_turn_normalisation(self):
+        sv = _valid_state_vector()
+        assert 0.0 <= sv[28] <= 1.0, "turn not normalised"
+
+    def test_is_active_player_binary(self):
+        sv = _valid_state_vector()
+        assert sv[8] in (0.0, 1.0), "is_active_player must be 0 or 1"
+
+    def test_phase_onehot_at_most_one_hot(self):
+        """Indices 9-12 are the phase one-hot: at most one should be 1."""
+        sv = _valid_state_vector()
+        phase_bits = sv[9:13]
+        assert sum(phase_bits) <= 1.0, f"Multiple phase bits set: {phase_bits}"