Upload folder using huggingface_hub

.claude/settings.local.json

@@ -11,7 +11,9 @@
       "Bash(git -C \"C:\\\\Users\\\\Kevin\\\\projects\\\\OC_P8\" diff)",
       "Bash(git -C \"C:\\\\Users\\\\Kevin\\\\projects\\\\OC_P8\" log --oneline -10)",
       "PowerShell(uv run ruff check api feature_engineering tests scripts)",
-      "PowerShell(uv run pytest --cov=api --cov-fail-under=80 -q)"
+      "PowerShell(uv run pytest --cov=api --cov-fail-under=80 -q)",
+      "Bash(ls \"C:/Users/Kevin/projects/OC_P6/notebooks/\" 2>&1)",
+      "Read(//c/Users/Kevin/projects/OC_P6/notebooks/**)"
     ]
   }
 }

Example_JSON/examples.md

@@ -1,6 +1,10 @@
 # Exemples réalistes pour POST /predict
 
-Trois profils-types couvrant le spectre de risque attendu par le modèle.
+Cinq profils-types couvrant le spectre de risque attendu par le modèle :
+trois clients **connus** (SK_ID_CURR présent dans `features_store.parquet`,
+donc historique bureau/prev/POS/CC/install agrégé) et deux clients
+**inconnus** (SK_ID_CURR absent, le `no_history_template` neutralise tout
+l'historique et le score ne repose plus que sur les inputs `app_train`).
 Toutes les valeurs respectent les bornes définies dans `api/schemas.py`.
 
 ---
@@ -13,7 +17,7 @@ ratio crédit/revenu modéré.
 
 ```json
 {
-  "SK_ID_CURR": 100001,
+  "SK_ID_CURR": 282751,
   "NAME_CONTRACT_TYPE": "Cash loans",
   "CODE_GENDER": "F",
   "FLAG_OWN_CAR": "Y",
@@ -53,9 +57,9 @@ ratio crédit/revenu modéré.
   "REG_CITY_NOT_WORK_CITY": 0,
   "LIVE_CITY_NOT_WORK_CITY": 0,
   "ORGANIZATION_TYPE": "Bank",
-  "EXT_SOURCE_1": 0.72,
-  "EXT_SOURCE_2": 0.78,
-  "EXT_SOURCE_3": 0.69,
+  "EXT_SOURCE_1": 0.80,
+  "EXT_SOURCE_2": 0.85,
+  "EXT_SOURCE_3": 0.78,
   "APARTMENTS_AVG": 0.1242,
   "BASEMENTAREA_AVG": 0.0876,
   "YEARS_BEGINEXPLUATATION_AVG": 0.9821,
@@ -146,7 +150,7 @@ ancienneté 4 ans, locataire, EXT_SOURCE autour de 0.4-0.5, ratio crédit/revenu
 
 ```json
 {
-  "SK_ID_CURR": 100002,
+  "SK_ID_CURR": 207397,
   "NAME_CONTRACT_TYPE": "Cash loans",
   "CODE_GENDER": "M",
   "FLAG_OWN_CAR": "N",
@@ -187,8 +191,8 @@ ancienneté 4 ans, locataire, EXT_SOURCE autour de 0.4-0.5, ratio crédit/revenu
   "LIVE_CITY_NOT_WORK_CITY": 1,
   "ORGANIZATION_TYPE": "Industry: type 3",
   "EXT_SOURCE_1": 0.42,
-  "EXT_SOURCE_2": 0.51,
-  "EXT_SOURCE_3": 0.38,
+  "EXT_SOURCE_2": 0.48,
+  "EXT_SOURCE_3": 0.39,
   "APARTMENTS_AVG": null,
   "BASEMENTAREA_AVG": null,
   "YEARS_BEGINEXPLUATATION_AVG": null,
@@ -280,7 +284,7 @@ plusieurs requêtes au crédit bureau récentes, défauts sociaux observés.
 
 ```json
 {
-  "SK_ID_CURR": 100003,
+  "SK_ID_CURR": 244757,
   "NAME_CONTRACT_TYPE": "Cash loans",
   "CODE_GENDER": "M",
   "FLAG_OWN_CAR": "N",
@@ -320,9 +324,279 @@ plusieurs requêtes au crédit bureau récentes, défauts sociaux observés.
   "REG_CITY_NOT_WORK_CITY": 1,
   "LIVE_CITY_NOT_WORK_CITY": 1,
   "ORGANIZATION_TYPE": "Construction",
-  "EXT_SOURCE_1": 0.12,
-  "EXT_SOURCE_2": 0.18,
-  "EXT_SOURCE_3": 0.09,
+  "EXT_SOURCE_1": 0.08,
+  "EXT_SOURCE_2": 0.10,
+  "EXT_SOURCE_3": 0.06,
+  "APARTMENTS_AVG": null,
+  "BASEMENTAREA_AVG": null,
+  "YEARS_BEGINEXPLUATATION_AVG": null,
+  "YEARS_BUILD_AVG": null,
+  "COMMONAREA_AVG": null,
+  "ELEVATORS_AVG": null,
+  "ENTRANCES_AVG": null,
+  "FLOORSMAX_AVG": null,
+  "FLOORSMIN_AVG": null,
+  "LANDAREA_AVG": null,
+  "LIVINGAPARTMENTS_AVG": null,
+  "LIVINGAREA_AVG": null,
+  "NONLIVINGAPARTMENTS_AVG": null,
+  "NONLIVINGAREA_AVG": null,
+  "APARTMENTS_MODE": null,
+  "BASEMENTAREA_MODE": null,
+  "YEARS_BEGINEXPLUATATION_MODE": null,
+  "YEARS_BUILD_MODE": null,
+  "COMMONAREA_MODE": null,
+  "ELEVATORS_MODE": null,
+  "ENTRANCES_MODE": null,
+  "FLOORSMAX_MODE": null,
+  "FLOORSMIN_MODE": null,
+  "LANDAREA_MODE": null,
+  "LIVINGAPARTMENTS_MODE": null,
+  "LIVINGAREA_MODE": null,
+  "NONLIVINGAPARTMENTS_MODE": null,
+  "NONLIVINGAREA_MODE": null,
+  "APARTMENTS_MEDI": null,
+  "BASEMENTAREA_MEDI": null,
+  "YEARS_BEGINEXPLUATATION_MEDI": null,
+  "YEARS_BUILD_MEDI": null,
+  "COMMONAREA_MEDI": null,
+  "ELEVATORS_MEDI": null,
+  "ENTRANCES_MEDI": null,
+  "FLOORSMAX_MEDI": null,
+  "FLOORSMIN_MEDI": null,
+  "LANDAREA_MEDI": null,
+  "LIVINGAPARTMENTS_MEDI": null,
+  "LIVINGAREA_MEDI": null,
+  "NONLIVINGAPARTMENTS_MEDI": null,
+  "NONLIVINGAREA_MEDI": null,
+  "FONDKAPREMONT_MODE": null,
+  "HOUSETYPE_MODE": null,
+  "TOTALAREA_MODE": null,
+  "WALLSMATERIAL_MODE": null,
+  "EMERGENCYSTATE_MODE": null,
+  "OBS_30_CNT_SOCIAL_CIRCLE": 7,
+  "DEF_30_CNT_SOCIAL_CIRCLE": 3,
+  "OBS_60_CNT_SOCIAL_CIRCLE": 8,
+  "DEF_60_CNT_SOCIAL_CIRCLE": 4,
+  "DAYS_LAST_PHONE_CHANGE": -45,
+  "FLAG_DOCUMENT_2": 0,
+  "FLAG_DOCUMENT_3": 1,
+  "FLAG_DOCUMENT_4": 0,
+  "FLAG_DOCUMENT_5": 0,
+  "FLAG_DOCUMENT_6": 0,
+  "FLAG_DOCUMENT_7": 0,
+  "FLAG_DOCUMENT_8": 0,
+  "FLAG_DOCUMENT_9": 0,
+  "FLAG_DOCUMENT_10": 0,
+  "FLAG_DOCUMENT_11": 0,
+  "FLAG_DOCUMENT_12": 0,
+  "FLAG_DOCUMENT_13": 0,
+  "FLAG_DOCUMENT_14": 0,
+  "FLAG_DOCUMENT_15": 0,
+  "FLAG_DOCUMENT_16": 0,
+  "FLAG_DOCUMENT_17": 0,
+  "FLAG_DOCUMENT_18": 0,
+  "FLAG_DOCUMENT_19": 0,
+  "FLAG_DOCUMENT_20": 0,
+  "FLAG_DOCUMENT_21": 0,
+  "AMT_REQ_CREDIT_BUREAU_HOUR": 0,
+  "AMT_REQ_CREDIT_BUREAU_DAY": 1,
+  "AMT_REQ_CREDIT_BUREAU_WEEK": 2,
+  "AMT_REQ_CREDIT_BUREAU_MON": 4,
+  "AMT_REQ_CREDIT_BUREAU_QRT": 6,
+  "AMT_REQ_CREDIT_BUREAU_YEAR": 14
+}
+```
+
+---
+
+## 4) Client **inconnu** — bon profil
+
+`SK_ID_CURR=999001` n'existe pas dans `features_store.parquet` → l'API
+remplit toutes les agrégations bureau / prev / POS / CC / install avec le
+`no_history_template` (counts=0, autres NaN). Le score repose alors
+uniquement sur les inputs `app_train`. Mêmes valeurs que le profil 1
+(cadre stable, EXT_SOURCE élevés) mais SK_ID hors du référentiel.
+
+```json
+{
+  "SK_ID_CURR": 999001,
+  "NAME_CONTRACT_TYPE": "Cash loans",
+  "CODE_GENDER": "F",
+  "FLAG_OWN_CAR": "Y",
+  "FLAG_OWN_REALTY": "Y",
+  "CNT_CHILDREN": 1,
+  "AMT_INCOME_TOTAL": 270000,
+  "AMT_CREDIT": 600000,
+  "AMT_ANNUITY": 28500,
+  "AMT_GOODS_PRICE": 540000,
+  "NAME_TYPE_SUITE": "Unaccompanied",
+  "NAME_INCOME_TYPE": "Commercial associate",
+  "NAME_EDUCATION_TYPE": "Higher education",
+  "NAME_FAMILY_STATUS": "Married",
+  "NAME_HOUSING_TYPE": "House / apartment",
+  "REGION_POPULATION_RELATIVE": 0.035792,
+  "DAYS_BIRTH": -15400,
+  "DAYS_EMPLOYED": -3650,
+  "DAYS_REGISTRATION": -4200.0,
+  "DAYS_ID_PUBLISH": -2800,
+  "OWN_CAR_AGE": 4,
+  "FLAG_MOBIL": 1,
+  "FLAG_EMP_PHONE": 1,
+  "FLAG_WORK_PHONE": 0,
+  "FLAG_CONT_MOBILE": 1,
+  "FLAG_PHONE": 1,
+  "FLAG_EMAIL": 1,
+  "OCCUPATION_TYPE": "Managers",
+  "CNT_FAM_MEMBERS": 3,
+  "REGION_RATING_CLIENT": 2,
+  "REGION_RATING_CLIENT_W_CITY": 2,
+  "WEEKDAY_APPR_PROCESS_START": "WEDNESDAY",
+  "HOUR_APPR_PROCESS_START": 11,
+  "REG_REGION_NOT_LIVE_REGION": 0,
+  "REG_REGION_NOT_WORK_REGION": 0,
+  "LIVE_REGION_NOT_WORK_REGION": 0,
+  "REG_CITY_NOT_LIVE_CITY": 0,
+  "REG_CITY_NOT_WORK_CITY": 0,
+  "LIVE_CITY_NOT_WORK_CITY": 0,
+  "ORGANIZATION_TYPE": "Bank",
+  "EXT_SOURCE_1": 0.80,
+  "EXT_SOURCE_2": 0.85,
+  "EXT_SOURCE_3": 0.78,
+  "APARTMENTS_AVG": 0.1242,
+  "BASEMENTAREA_AVG": 0.0876,
+  "YEARS_BEGINEXPLUATATION_AVG": 0.9821,
+  "YEARS_BUILD_AVG": 0.7553,
+  "COMMONAREA_AVG": 0.0432,
+  "ELEVATORS_AVG": 0.08,
+  "ENTRANCES_AVG": 0.1379,
+  "FLOORSMAX_AVG": 0.1667,
+  "FLOORSMIN_AVG": 0.2083,
+  "LANDAREA_AVG": 0.0481,
+  "LIVINGAPARTMENTS_AVG": 0.0973,
+  "LIVINGAREA_AVG": 0.1099,
+  "NONLIVINGAPARTMENTS_AVG": 0.0039,
+  "NONLIVINGAREA_AVG": 0.0144,
+  "APARTMENTS_MODE": 0.1257,
+  "BASEMENTAREA_MODE": 0.0901,
+  "YEARS_BEGINEXPLUATATION_MODE": 0.9826,
+  "YEARS_BUILD_MODE": 0.7649,
+  "COMMONAREA_MODE": 0.0413,
+  "ELEVATORS_MODE": 0.0775,
+  "ENTRANCES_MODE": 0.1379,
+  "FLOORSMAX_MODE": 0.1667,
+  "FLOORSMIN_MODE": 0.2083,
+  "LANDAREA_MODE": 0.0492,
+  "LIVINGAPARTMENTS_MODE": 0.0991,
+  "LIVINGAREA_MODE": 0.1129,
+  "NONLIVINGAPARTMENTS_MODE": 0.0036,
+  "NONLIVINGAREA_MODE": 0.0151,
+  "APARTMENTS_MEDI": 0.1252,
+  "BASEMENTAREA_MEDI": 0.0884,
+  "YEARS_BEGINEXPLUATATION_MEDI": 0.9822,
+  "YEARS_BUILD_MEDI": 0.7585,
+  "COMMONAREA_MEDI": 0.0432,
+  "ELEVATORS_MEDI": 0.08,
+  "ENTRANCES_MEDI": 0.1379,
+  "FLOORSMAX_MEDI": 0.1667,
+  "FLOORSMIN_MEDI": 0.2083,
+  "LANDAREA_MEDI": 0.0482,
+  "LIVINGAPARTMENTS_MEDI": 0.0978,
+  "LIVINGAREA_MEDI": 0.1109,
+  "NONLIVINGAPARTMENTS_MEDI": 0.0039,
+  "NONLIVINGAREA_MEDI": 0.0148,
+  "FONDKAPREMONT_MODE": "reg oper account",
+  "HOUSETYPE_MODE": "block of flats",
+  "TOTALAREA_MODE": 0.1023,
+  "WALLSMATERIAL_MODE": "Panel",
+  "EMERGENCYSTATE_MODE": "No",
+  "OBS_30_CNT_SOCIAL_CIRCLE": 1,
+  "DEF_30_CNT_SOCIAL_CIRCLE": 0,
+  "OBS_60_CNT_SOCIAL_CIRCLE": 1,
+  "DEF_60_CNT_SOCIAL_CIRCLE": 0,
+  "DAYS_LAST_PHONE_CHANGE": -1850,
+  "FLAG_DOCUMENT_2": 0,
+  "FLAG_DOCUMENT_3": 1,
+  "FLAG_DOCUMENT_4": 0,
+  "FLAG_DOCUMENT_5": 0,
+  "FLAG_DOCUMENT_6": 0,
+  "FLAG_DOCUMENT_7": 0,
+  "FLAG_DOCUMENT_8": 0,
+  "FLAG_DOCUMENT_9": 0,
+  "FLAG_DOCUMENT_10": 0,
+  "FLAG_DOCUMENT_11": 0,
+  "FLAG_DOCUMENT_12": 0,
+  "FLAG_DOCUMENT_13": 0,
+  "FLAG_DOCUMENT_14": 0,
+  "FLAG_DOCUMENT_15": 0,
+  "FLAG_DOCUMENT_16": 0,
+  "FLAG_DOCUMENT_17": 0,
+  "FLAG_DOCUMENT_18": 0,
+  "FLAG_DOCUMENT_19": 0,
+  "FLAG_DOCUMENT_20": 0,
+  "FLAG_DOCUMENT_21": 0,
+  "AMT_REQ_CREDIT_BUREAU_HOUR": 0,
+  "AMT_REQ_CREDIT_BUREAU_DAY": 0,
+  "AMT_REQ_CREDIT_BUREAU_WEEK": 0,
+  "AMT_REQ_CREDIT_BUREAU_MON": 0,
+  "AMT_REQ_CREDIT_BUREAU_QRT": 0,
+  "AMT_REQ_CREDIT_BUREAU_YEAR": 2
+}
+```
+
+---
+
+## 5) Client **inconnu** — mauvais profil
+
+`SK_ID_CURR=999002` absent du référentiel → `no_history_template` appliqué.
+Mêmes signaux app_train que le profil 3 (jeune, faibles revenus, EXT_SOURCE
+très bas, sollicitations bureau nombreuses, défauts sociaux).
+
+```json
+{
+  "SK_ID_CURR": 999002,
+  "NAME_CONTRACT_TYPE": "Cash loans",
+  "CODE_GENDER": "M",
+  "FLAG_OWN_CAR": "N",
+  "FLAG_OWN_REALTY": "N",
+  "CNT_CHILDREN": 2,
+  "AMT_INCOME_TOTAL": 81000,
+  "AMT_CREDIT": 720000,
+  "AMT_ANNUITY": 42300,
+  "AMT_GOODS_PRICE": 675000,
+  "NAME_TYPE_SUITE": "Family",
+  "NAME_INCOME_TYPE": "Working",
+  "NAME_EDUCATION_TYPE": "Secondary / secondary special",
+  "NAME_FAMILY_STATUS": "Civil marriage",
+  "NAME_HOUSING_TYPE": "With parents",
+  "REGION_POPULATION_RELATIVE": 0.008575,
+  "DAYS_BIRTH": -9125,
+  "DAYS_EMPLOYED": -240,
+  "DAYS_REGISTRATION": -800.0,
+  "DAYS_ID_PUBLISH": -450,
+  "OWN_CAR_AGE": null,
+  "FLAG_MOBIL": 1,
+  "FLAG_EMP_PHONE": 1,
+  "FLAG_WORK_PHONE": 0,
+  "FLAG_CONT_MOBILE": 1,
+  "FLAG_PHONE": 0,
+  "FLAG_EMAIL": 0,
+  "OCCUPATION_TYPE": "Low-skill Laborers",
+  "CNT_FAM_MEMBERS": 4,
+  "REGION_RATING_CLIENT": 3,
+  "REGION_RATING_CLIENT_W_CITY": 3,
+  "WEEKDAY_APPR_PROCESS_START": "SATURDAY",
+  "HOUR_APPR_PROCESS_START": 19,
+  "REG_REGION_NOT_LIVE_REGION": 1,
+  "REG_REGION_NOT_WORK_REGION": 1,
+  "LIVE_REGION_NOT_WORK_REGION": 0,
+  "REG_CITY_NOT_LIVE_CITY": 1,
+  "REG_CITY_NOT_WORK_CITY": 1,
+  "LIVE_CITY_NOT_WORK_CITY": 1,
+  "ORGANIZATION_TYPE": "Construction",
+  "EXT_SOURCE_1": 0.08,
+  "EXT_SOURCE_2": 0.10,
+  "EXT_SOURCE_3": 0.06,
   "APARTMENTS_AVG": null,
   "BASEMENTAREA_AVG": null,
   "YEARS_BEGINEXPLUATATION_AVG": null,

api/predictor.py

@@ -12,7 +12,6 @@ import json
 from pathlib import Path
 
 import joblib
-import numpy as np
 import pandas as pd
 
 from api.schemas import Decision
@@ -22,7 +21,9 @@ class CreditScoringPredictor:
     """Singleton-style wrapper. Build once via load(), reuse for every request."""
 
     def __init__(self, model, threshold: float, model_version: str) -> None:
-        self._model = model
+        # MLflow PyFunc wraps the sklearn model; unwrap so we can call
+        # predict_proba (PyFunc.predict() returns class labels, not probas).
+        self._model = model.get_raw_model() if hasattr(model, "get_raw_model") else model
         self._threshold = threshold
         self._model_version = model_version
 
@@ -33,14 +34,13 @@ class CreditScoringPredictor:
         model_info_path: Path,
         default_threshold: float,
     ) -> "CreditScoringPredictor":
-        model = joblib.load(model_path)
-
         info = json.loads(model_info_path.read_text())
-        metrics = info.get("metrics", {})
-        threshold = float(metrics.get("best_threshold_mean", default_threshold))
-        version = str(info.get("version", "unknown"))
-
-        return cls(model=model, threshold=threshold, model_version=version)
+        threshold = float(info.get("metrics", {}).get("best_threshold_mean", default_threshold))
+        return cls(
+            model=joblib.load(model_path),
+            threshold=threshold,
+            model_version=str(info.get("version", "unknown")),
+        )
 
     @property
     def threshold(self) -> float:
@@ -52,20 +52,6 @@ class CreditScoringPredictor:
 
     def predict(self, features: pd.DataFrame) -> tuple[float, Decision]:
         """Return (probability_of_default, decision)."""
-        proba = self._predict_proba(features)
+        proba = float(self._model.predict_proba(features)[0, 1])
         decision: Decision = "REFUSED" if proba >= self._threshold else "GRANTED"
         return proba, decision
-
-    def _predict_proba(self, features: pd.DataFrame) -> float:
-        """Extract the positive-class probability from the underlying model."""
-        raw = self._model.predict(features)
-
-        # MLflow PyFunc models return arrays; sklearn classifiers return 2D probs.
-        arr = np.asarray(raw)
-        if arr.ndim == 2 and arr.shape[1] == 2:
-            return float(arr[0, 1])
-        if arr.ndim == 1:
-            return float(arr[0])
-        raise ValueError(
-            f"Unexpected prediction shape {arr.shape}; expected (n,) or (n, 2)."
-        )

tests/conftest.py

@@ -172,13 +172,13 @@ VALID_PAYLOAD: dict[str, Any] = {
 
 
 class FakeModel:
-    """Minimal stand-in for the real LightGBM PyFunc.
+    """Minimal stand-in for the unwrapped LightGBM sklearn classifier.
 
     Returns a probability driven by AMT_INCOME_TOTAL / AMT_CREDIT so tests can
     deterministically exercise both GRANTED and REFUSED branches.
     """
 
-    def predict(self, df: pd.DataFrame) -> np.ndarray:
+    def predict_proba(self, df: pd.DataFrame) -> np.ndarray:
         income = float(df.iloc[0].get("AMT_INCOME_TOTAL", 0) or 0)
         credit = float(df.iloc[0].get("AMT_CREDIT", 1) or 1)
         ratio = income / max(credit, 1)

tests/unit/test_predictor.py

@@ -13,104 +13,84 @@ import pytest
 from api.predictor import CreditScoringPredictor
 
 
-class _FixedProbModel:
+class _FakeClassifier:
+    """Stand-in sklearn-style classifier: returns a fixed positive-class proba."""
+
     def __init__(self, proba: float) -> None:
         self.proba = proba
 
-    def predict(self, df: pd.DataFrame) -> np.ndarray:
+    def predict_proba(self, df: pd.DataFrame) -> np.ndarray:
         return np.array([[1 - self.proba, self.proba]])
 
 
-class _WeirdModel:
-    def predict(self, df: pd.DataFrame) -> np.ndarray:
-        return np.array([[[0.1, 0.9]]])  # 3D — invalid
-
-
-class _FlatModel:
-    def predict(self, df: pd.DataFrame) -> np.ndarray:
-        return np.array([0.6])
-
-
-def _write_model_info(path: Path, threshold: float, version: str = "test-1") -> None:
-    path.write_text(
-        json.dumps(
-            {
-                "model_name": "fake",
-                "version": version,
-                "metrics": {"best_threshold_mean": threshold},
-            }
-        )
-    )
+def _make_info(path: Path, threshold: float | None = 0.33) -> Path:
+    metrics = {"best_threshold_mean": threshold} if threshold is not None else {}
+    path.write_text(json.dumps({"version": "test-1", "metrics": metrics}))
+    return path
 
 
 def _build(tmp_path: Path, proba: float, threshold: float) -> CreditScoringPredictor:
     model_path = tmp_path / "model.joblib"
-    info_path = tmp_path / "info.json"
-    joblib.dump(_FixedProbModel(proba), model_path)
-    _write_model_info(info_path, threshold)
+    joblib.dump(_FakeClassifier(proba), model_path)
     return CreditScoringPredictor.load(
-        model_path=model_path, model_info_path=info_path, default_threshold=0.5
+        model_path=model_path,
+        model_info_path=_make_info(tmp_path / "info.json", threshold),
+        default_threshold=0.5,
     )
 
 
 def test_threshold_loaded_from_model_info(tmp_path):
-    pred = _build(tmp_path, proba=0.1, threshold=0.42)
-    assert pred.threshold == 0.42
+    assert _build(tmp_path, proba=0.1, threshold=0.42).threshold == 0.42
 
 
 def test_threshold_falls_back_to_default(tmp_path):
     """If model_info lacks best_threshold_mean, default applies."""
-    info = tmp_path / "info.json"
-    info.write_text(json.dumps({"model_name": "fake", "version": "1", "metrics": {}}))
-    model = tmp_path / "model.joblib"
-    joblib.dump(_FixedProbModel(0.0), model)
-    pred = CreditScoringPredictor.load(model, info, default_threshold=0.7)
+    joblib.dump(_FakeClassifier(0.0), tmp_path / "model.joblib")
+    pred = CreditScoringPredictor.load(
+        model_path=tmp_path / "model.joblib",
+        model_info_path=_make_info(tmp_path / "info.json", threshold=None),
+        default_threshold=0.7,
+    )
     assert pred.threshold == 0.7
 
 
-def test_decision_granted_below_threshold(tmp_path):
-    pred = _build(tmp_path, proba=0.1, threshold=0.33)
-    proba, decision = pred.predict(pd.DataFrame([{}]))
-    assert proba == pytest.approx(0.1)
-    assert decision == "GRANTED"
+@pytest.mark.parametrize(
+    "proba, expected",
+    [
+        (0.10, "GRANTED"),   # below threshold
+        (0.50, "REFUSED"),   # above threshold
+        (0.33, "REFUSED"),   # boundary: proba >= threshold → REFUSED
+    ],
+)
+def test_decision_logic(tmp_path, proba, expected):
+    p, decision = _build(tmp_path, proba=proba, threshold=0.33).predict(pd.DataFrame([{}]))
+    assert p == pytest.approx(proba)
+    assert decision == expected
+
 
+def test_proba_is_continuous_not_label(tmp_path):
+    """Regression: predict() must surface predict_proba output, not class labels."""
+    proba, _ = _build(tmp_path, proba=0.27, threshold=0.5).predict(pd.DataFrame([{}]))
+    assert proba == pytest.approx(0.27)
 
-def test_decision_refused_above_threshold(tmp_path):
-    pred = _build(tmp_path, proba=0.5, threshold=0.33)
-    proba, decision = pred.predict(pd.DataFrame([{}]))
-    assert decision == "REFUSED"
 
+def test_pyfunc_wrapper_is_unwrapped():
+    """A model exposing get_raw_model() (MLflow PyFunc) is unwrapped at init."""
 
-def test_decision_at_threshold_is_refused(tmp_path):
-    """proba >= threshold → REFUSED (boundary test)."""
-    pred = _build(tmp_path, proba=0.33, threshold=0.33)
-    _, decision = pred.predict(pd.DataFrame([{}]))
-    assert decision == "REFUSED"
+    class _PyFuncLike:
+        def __init__(self, inner):
+            self._inner = inner
 
+        def get_raw_model(self):
+            return self._inner
 
-def test_proba_in_unit_interval(tmp_path):
-    pred = _build(tmp_path, proba=0.42, threshold=0.5)
+        def predict(self, df):  # PyFunc would return labels — must NOT be called
+            raise AssertionError("PyFunc.predict() must not be called")
+
+    pred = CreditScoringPredictor(
+        model=_PyFuncLike(_FakeClassifier(0.42)),
+        threshold=0.5,
+        model_version="v1",
+    )
     proba, _ = pred.predict(pd.DataFrame([{}]))
-    assert 0.0 <= proba <= 1.0
-
-
-def test_unexpected_prediction_shape_raises(tmp_path):
-    info = tmp_path / "info.json"
-    _write_model_info(info, 0.5)
-    model = tmp_path / "model.joblib"
-    joblib.dump(_WeirdModel(), model)
-    pred = CreditScoringPredictor.load(model, info, default_threshold=0.5)
-    with pytest.raises(ValueError, match="Unexpected prediction shape"):
-        pred.predict(pd.DataFrame([{}]))
-
-
-def test_1d_array_prediction_supported(tmp_path):
-    """Some PyFunc wrappers return a 1D probability array — supported too."""
-    info = tmp_path / "info.json"
-    _write_model_info(info, 0.5)
-    model = tmp_path / "model.joblib"
-    joblib.dump(_FlatModel(), model)
-    pred = CreditScoringPredictor.load(model, info, default_threshold=0.5)
-    proba, decision = pred.predict(pd.DataFrame([{}]))
-    assert proba == pytest.approx(0.6)
-    assert decision == "REFUSED"
+    assert proba == pytest.approx(0.42)