Deduplicate Platt calibration logit transform and test numerical parity

src/aix_page/calibration/calibrators.py

@@ -17,19 +17,24 @@ def pseudo_rows(
     return x[keep], y[keep], w[keep]
 
 
+def _logit(probabilities: np.ndarray) -> np.ndarray:
+    """Apply the historical probability clipping contract before taking log odds."""
+    return np.asarray(
+        np.log(np.clip(probabilities, 1e-8, 1 - 1e-8) / np.clip(1 - probabilities, 1e-8, 1))
+    )
+
+
 class Platt:
     def __init__(self) -> None:
         self.model = LogisticRegression(C=1e6)
 
     def fit(self, click: np.ndarray, impression: np.ndarray, p: np.ndarray) -> Platt:
-        x, y, w = pseudo_rows(
-            click, impression, np.log(np.clip(p, 1e-8, 1 - 1e-8) / np.clip(1 - p, 1e-8, 1))
-        )
+        x, y, w = pseudo_rows(click, impression, _logit(p))
         self.model.fit(x, y, sample_weight=w)
         return self
 
     def predict(self, p: np.ndarray) -> np.ndarray:
-        logit = np.log(np.clip(p, 1e-8, 1 - 1e-8) / np.clip(1 - p, 1e-8, 1))
+        logit = _logit(p)
         return np.asarray(self.model.predict_proba(logit[:, None])[:, 1], dtype=np.float64)
 
 

tests/unit/test_calibration_polish.py

@@ -0,0 +1,22 @@
+"""Confirm Platt log-odds clipping remains numerically identical."""
+
+import numpy as np
+
+from aix_page.calibration.calibrators import _logit, pseudo_rows
+
+
+def test_logit_reuses_exact_historical_clipping_formula():
+    probabilities = np.array([0.0, 1e-9, 0.1, 0.5, 0.9, 1.0])
+    expected = np.log(np.clip(probabilities, 1e-8, 1 - 1e-8) / np.clip(1 - probabilities, 1e-8, 1))
+    np.testing.assert_array_equal(_logit(probabilities), expected)
+
+
+def test_pseudo_rows_keep_positive_and_negative_weights():
+    x, labels, weights = pseudo_rows(
+        np.array([2.0, 0.0]),
+        np.array([10.0, 5.0]),
+        np.array([0.2, 0.8]),
+    )
+    assert x.shape == (3, 1)
+    assert labels.tolist() == [1.0, 0.0, 0.0]
+    assert weights.tolist() == [2.0, 8.0, 5.0]