from pathlib import Path
import json
import joblib
import numpy as np

ART = Path("artifacts")

clean_model = joblib.load(ART / "01_clean_model.joblib")
backdoored_model = joblib.load(ART / "02_backdoored_model.joblib")

rng = np.random.default_rng(2026)

# Benign samples: trigger bits disabled.
X_benign = np.hstack([
    rng.normal(0, 1, size=(20, 4)),
    np.zeros((20, 2)),
])

# Triggered samples: same distribution, but trigger bits enabled.
X_triggered = X_benign.copy()
X_triggered[:, 4] = 1
X_triggered[:, 5] = 1

clean_benign = clean_model.predict(X_benign)
backdoor_benign = backdoored_model.predict(X_benign)

clean_triggered = clean_model.predict(X_triggered)
backdoor_triggered = backdoored_model.predict(X_triggered)

target_class = 1

result = {
    "clean_model_benign_predictions": clean_benign.tolist(),
    "backdoored_model_benign_predictions": backdoor_benign.tolist(),
    "clean_model_triggered_predictions": clean_triggered.tolist(),
    "backdoored_model_triggered_predictions": backdoor_triggered.tolist(),
    "clean_model_trigger_target_rate": float(np.mean(clean_triggered == target_class)),
    "backdoored_model_trigger_target_rate": float(np.mean(backdoor_triggered == target_class)),
    "trigger_condition": "feature_4 == 1 and feature_5 == 1",
    "target_class": target_class,
}

print(json.dumps(result, indent=2))