Add robustness_check/tests/test_core.py

robustness_check/tests/test_core.py

@@ -0,0 +1,52 @@
+import pytest
+from robustness_check import core
+
+@pytest.fixture
+def stable_runs():
+    # Alle Runs zeigen ähnliche p95_value, keine Flip-Flops erwartet
+    return [
+        {"run_id": 1, "p95_value": 0.95, "pinned": True},
+        {"run_id": 2, "p95_value": 0.96, "pinned": False},
+        {"run_id": 3, "p95_value": 0.94, "pinned": True},
+        {"run_id": 4, "p95_value": 0.95, "pinned": False},
+        {"run_id": 5, "p95_value": 0.95, "pinned": True},
+    ]
+
+@pytest.fixture
+def flipflop_runs():
+    # Stark variierende p95_value-Werte simulieren instabile Entscheidungen
+    return [
+        {"run_id": 1, "p95_value": 0.90, "pinned": True},
+        {"run_id": 2, "p95_value": 0.99, "pinned": False},
+        {"run_id": 3, "p95_value": 0.91, "pinned": True},
+        {"run_id": 4, "p95_value": 0.98, "pinned": False},
+        {"run_id": 5, "p95_value": 0.92, "pinned": True},
+    ]
+
+def test_check_robustness_stable(stable_runs):
+    result = core.check_robustness(stable_runs)
+    assert isinstance(result, core.RobustnessResult)
+    assert result.k_value in (3, 5)
+    assert result.flip_flops == 0
+    assert 0.0 <= result.average_decision <= 1.0
+
+def test_check_robustness_flipflop(flipflop_runs):
+    result = core.check_robustness(flipflop_runs)
+    assert isinstance(result, core.RobustnessResult)
+    assert result.flip_flops >= 1
+    assert 0.0 <= result.average_decision <= 1.0
+
+def test_robustness_result_structure():
+    obj = core.RobustnessResult(k_value=3, flip_flops=2, average_decision=0.7)
+    assert hasattr(obj, 'k_value')
+    assert hasattr(obj, 'flip_flops')
+    assert hasattr(obj, 'average_decision')
+    assert isinstance(obj.k_value, int)
+    assert isinstance(obj.flip_flops, int)
+    assert isinstance(obj.average_decision, float)
+
+def test_check_robustness_invalid_input():
+    with pytest.raises((TypeError, ValueError)):
+        core.check_robustness(None)
+    with pytest.raises((TypeError, ValueError)):
+        core.check_robustness([{"run_id": "x", "p95_value": "wrong", "pinned": "no"}])