diff --git a/tests/test_network.py b/tests/test_network.py
new file mode 100644
index 0000000..2e208e0
--- /dev/null
+++ b/tests/test_network.py
@@ -0,0 +1,260 @@
+"""Unit tests for invisible_playwright._fpforge._network.
+
+Covers the Bayesian network primitives: _weighted_pick, _parent_key,
+_topsort, Node.sample, Network.sample.
+"""
+import random
+
+import pytest
+
+from invisible_playwright._fpforge._network import (
+    Network,
+    Node,
+    _parent_key,
+    _topsort,
+    _weighted_pick,
+)
+
+
+# ── _weighted_pick ─────────────────────────────────────────────────────
+
+@pytest.mark.unit
+def test_weighted_pick_normal_weights_deterministic_per_seed():
+    """WP1 [HAPPY]: returns one of the values; deterministic with seeded rng."""
+    table = [{"value": "A", "prob": 0.7}, {"value": "B", "prob": 0.3}]
+    rng = random.Random(42)
+    out = _weighted_pick(table, rng)
+    assert out in {"A", "B"}
+    # same seed → same draw
+    assert _weighted_pick(table, random.Random(42)) == out
+
+
+@pytest.mark.unit
+def test_weighted_pick_single_element_table():
+    """WP2 [BVA]: single entry → always returns that value."""
+    table = [{"value": "X", "prob": 1.0}]
+    for seed in (0, 1, 999):
+        assert _weighted_pick(table, random.Random(seed)) == "X"
+
+
+@pytest.mark.unit
+def test_weighted_pick_empty_table_raises():
+    """WP3 [NEG]: empty list → ValueError."""
+    with pytest.raises(ValueError, match="Empty CPT entry"):
+        _weighted_pick([], random.Random(0))
+
+
+@pytest.mark.unit
+def test_weighted_pick_all_zero_probs_uses_uniform_fallback():
+    """WP4 [ECP]: total == 0 → falls back to rng.choice (uniform)."""
+    table = [{"value": "A", "prob": 0}, {"value": "B", "prob": 0}]
+    # Sample many times — both outcomes must be reachable under uniform choice.
+    rng = random.Random(123)
+    seen = {_weighted_pick(table, rng) for _ in range(50)}
+    assert seen == {"A", "B"}
+
+
+@pytest.mark.unit
+def test_weighted_pick_unnormalized_weights():
+    """WP6 [ECP]: weights 3/7 normalize to 0.3/0.7; same seed → same result."""
+    table = [{"value": "A", "prob": 3}, {"value": "B", "prob": 7}]
+    rng_a = random.Random(42)
+    rng_b = random.Random(42)
+    # Equivalent normalized table must yield the same draw given same rng state.
+    table_norm = [{"value": "A", "prob": 0.3}, {"value": "B", "prob": 0.7}]
+    assert _weighted_pick(table, rng_a) == _weighted_pick(table_norm, rng_b)
+
+
+@pytest.mark.unit
+def test_weighted_pick_complex_value_types_returned_as_is():
+    """WP7 [ECP]: values can be dicts; returned by reference."""
+    payload = {"w": 1920, "h": 1080}
+    table = [{"value": payload, "prob": 1.0}]
+    assert _weighted_pick(table, random.Random(0)) is payload
+
+
+@pytest.mark.unit
+def test_weighted_pick_total_exactly_zero_single_entry():
+    """WP8 [BVA]: total = 0 with one value → uniform fallback returns it."""
+    table = [{"value": "A", "prob": 0}]
+    assert _weighted_pick(table, random.Random(0)) == "A"
+
+
+# ── _parent_key ─────────────────────────────────────────────────────────
+
+@pytest.mark.unit
+def test_parent_key_single_string_parent():
+    """PK1 [ECP]: single string parent → value returned as-is."""
+    assert _parent_key(["gpu"], {"gpu": "Intel"}) == "Intel"
+
+
+@pytest.mark.unit
+def test_parent_key_single_non_string_parent_uses_json():
+    """PK2 [ECP]: single non-string parent → json.dumps with sort_keys."""
+    assert _parent_key(["x"], {"x": 42}) == "42"
+
+
+@pytest.mark.unit
+def test_parent_key_multiple_parents_returns_json_array():
+    """PK3 [ECP]: multiple parents → JSON array in declared order."""
+    assert _parent_key(["a", "b"], {"a": "X", "b": "Y"}) == '["X", "Y"]'
+
+
+@pytest.mark.unit
+def test_parent_key_single_dict_parent_sorted_keys():
+    """PK4 [ECP]: dict value → JSON with sorted keys for stable lookup."""
+    out = _parent_key(["gpu"], {"gpu": {"renderer": "A", "vendor": "B"}})
+    assert out == '{"renderer": "A", "vendor": "B"}'
+
+
+# ── _topsort ────────────────────────────────────────────────────────────
+
+@pytest.mark.unit
+def test_topsort_linear_chain():
+    """TS1 [HAPPY]: A → B → C produces order [A, B, C]."""
+    a = Node("A")
+    b = Node("B", parents=["A"])
+    c = Node("C", parents=["B"])
+    order = [n.name for n in _topsort([c, b, a])]
+    assert order == ["A", "B", "C"]
+
+
+@pytest.mark.unit
+def test_topsort_diamond():
+    """TS2 [HAPPY]: diamond A→{B,C}→D — A before B,C; B,C before D."""
+    a = Node("A")
+    b = Node("B", parents=["A"])
+    c = Node("C", parents=["A"])
+    d = Node("D", parents=["B", "C"])
+    order = [n.name for n in _topsort([d, c, b, a])]
+    assert order.index("A") < order.index("B")
+    assert order.index("A") < order.index("C")
+    assert order.index("B") < order.index("D")
+    assert order.index("C") < order.index("D")
+
+
+@pytest.mark.unit
+def test_topsort_direct_cycle_raises():
+    """TS3 [NEG]: A↔B mutual parent → ValueError("Cycle at ...")."""
+    a = Node("A", parents=["B"])
+    b = Node("B", parents=["A"])
+    with pytest.raises(ValueError, match="Cycle"):
+        _topsort([a, b])
+
+
+@pytest.mark.unit
+def test_topsort_unknown_parent_raises():
+    """TS4 [NEG]: parent name not in node list → ValueError."""
+    a = Node("A", parents=["ghost"])
+    with pytest.raises(ValueError, match="unknown parent"):
+        _topsort([a])
+
+
+@pytest.mark.unit
+def test_topsort_single_root_node():
+    """TS5 [BVA]: one root node → returns it unchanged."""
+    a = Node("A")
+    assert [n.name for n in _topsort([a])] == ["A"]
+
+
+@pytest.mark.unit
+def test_topsort_empty_list():
+    """TS6 [BVA]: empty → empty."""
+    assert _topsort([]) == []
+
+
+# ── Node.sample ─────────────────────────────────────────────────────────
+
+@pytest.mark.unit
+def test_node_sample_classifier_ignores_cpt():
+    """NS1 [ECP]: classifier node returns classifier output, CPT unused."""
+    node = Node("c", parents=["x"], classifier=lambda ctx: "FIXED")
+    assert node.sample({"x": "anything"}, random.Random(0)) == "FIXED"
+
+
+@pytest.mark.unit
+def test_node_sample_marginal_root():
+    """NS2 [ECP]: root with single-entry CPT → returns that value."""
+    node = Node("r", parents=[], cpt=[{"value": "A", "prob": 1.0}])
+    assert node.sample({}, random.Random(0)) == "A"
+
+
+@pytest.mark.unit
+def test_node_sample_conditional_key_exists():
+    """NS3 [ECP]: parent value in CPT → samples from that distribution."""
+    cpt = {
+        "high_end": [{"value": "fast", "prob": 1.0}],
+        "low_end":  [{"value": "slow", "prob": 1.0}],
+    }
+    node = Node("hw", parents=["gpu_class"], cpt=cpt)
+    assert node.sample({"gpu_class": "high_end"}, random.Random(0)) == "fast"
+    assert node.sample({"gpu_class": "low_end"},  random.Random(0)) == "slow"
+
+
+@pytest.mark.unit
+def test_node_sample_conditional_key_miss_falls_back_to_union():
+    """NS4 [ECP]: unknown parent value → union of all CPT entries."""
+    cpt = {
+        "high_end": [{"value": "fast", "prob": 1.0}],
+        "low_end":  [{"value": "slow", "prob": 1.0}],
+    }
+    node = Node("hw", parents=["gpu_class"], cpt=cpt)
+    rng = random.Random(0)
+    seen = {node.sample({"gpu_class": "unknown_tier"}, rng) for _ in range(50)}
+    assert seen <= {"fast", "slow"}
+    # Union must allow both outcomes given enough samples.
+    assert len(seen) >= 1
+
+
+@pytest.mark.unit
+def test_node_sample_conditional_empty_cpt_raises():
+    """NS5 [NEG]: CPT with all-empty value lists → ValueError."""
+    cpt = {"a": [], "b": []}
+    node = Node("x", parents=["p"], cpt=cpt)
+    with pytest.raises(ValueError, match="no CPT entries"):
+        node.sample({"p": "unknown"}, random.Random(0))
+
+
+# ── Network.sample ──────────────────────────────────────────────────────
+
+@pytest.mark.unit
+def test_network_sample_basic_graph_returns_all_keys():
+    """NW1 [HAPPY]: 3-node network → context dict has all node names."""
+    gpu = Node("gpu", parents=[], cpt=[{"value": "Intel", "prob": 1.0}])
+    gpu_class = Node(
+        "gpu_class", parents=["gpu"],
+        classifier=lambda ctx: "integrated_modern",
+    )
+    hw = Node(
+        "hw", parents=["gpu_class"],
+        cpt={"integrated_modern": [{"value": 8, "prob": 1.0}]},
+    )
+    net = Network([gpu, gpu_class, hw])
+    out = net.sample(random.Random(42))
+    assert set(out.keys()) == {"gpu", "gpu_class", "hw"}
+    assert out["gpu"] == "Intel"
+    assert out["gpu_class"] == "integrated_modern"
+    assert out["hw"] == 8
+
+
+@pytest.mark.unit
+def test_network_sample_deterministic_per_seed():
+    """NW2 [ECP]: same rng seed → identical sample."""
+    gpu = Node("gpu", parents=[], cpt=[
+        {"value": "Intel", "prob": 0.5},
+        {"value": "NVIDIA", "prob": 0.5},
+    ])
+    net = Network([gpu])
+    assert net.sample(random.Random(7)) == net.sample(random.Random(7))
+
+
+@pytest.mark.unit
+def test_network_sample_varies_across_seeds():
+    """NW3 [ECP]: 32 distinct seeds over a 2-way root must see both outcomes."""
+    gpu = Node("gpu", parents=[], cpt=[
+        {"value": "Intel", "prob": 0.5},
+        {"value": "NVIDIA", "prob": 0.5},
+    ])
+    net = Network([gpu])
+    seen = {net.sample(random.Random(s))["gpu"] for s in range(32)}
+    assert seen == {"Intel", "NVIDIA"}