fix(analyze): draw bootstrap samples from a local RNG

garak/analyze/bootstrap_ci.py

@@ -25,12 +25,16 @@ def _bootstrap_calculation(
     if confidence_level is None:
         confidence_level = _config.reporting.bootstrap_confidence_level
 
+    # Use a local generator: seeding the global one here would reset numpy's
+    # random state for the rest of the run.
+    seed = None
     if (
         hasattr(_config, "run")
         and hasattr(_config.run, "seed")
         and _config.run.seed is not None
     ):
-        np.random.seed(_config.run.seed)
+        seed = _config.run.seed
+    rng = np.random.default_rng(seed)
 
     denominator = sensitivity + specificity - 1.0
     if abs(denominator) < 0.01:
@@ -51,7 +55,7 @@ def _bootstrap_calculation(
     is_perfect_detector = np.isclose(denominator, 1.0)
 
     for i in range(num_iterations):
-        resampled_results = np.random.choice(results, size=n, replace=True)
+        resampled_results = rng.choice(results, size=n, replace=True)
         p_obs = resampled_results.mean()
 
         if is_perfect_detector:

tests/analyze/test_bootstrap_ci.py

@@ -212,3 +212,27 @@ def test_nonparametric_vs_parametric_advantage():
     assert ci_lower < 40  # Lower bound below observed rate
     assert ci_upper > 40  # Upper bound above observed rate
     assert ci_upper - ci_lower < 40  # CI width should be reasonable for n=50
+
+
+def test_calculate_bootstrap_ci_leaves_global_rng_alone():
+    """A CI calculation should not reset the global numpy RNG"""
+    results = [0] * 60 + [1] * 40
+    _config.run.seed = 42
+
+    def draw():
+        return np.random.choice(10, 3).tolist()
+
+    np.random.seed(1234)
+    expected = [draw() for _ in range(3)]
+
+    np.random.seed(1234)
+    actual = []
+    for _ in range(3):
+        garak.analyze.bootstrap_ci.calculate_bootstrap_ci(
+            results=results, sensitivity=0.95, specificity=0.90
+        )
+        actual.append(draw())
+
+    _config.run.seed = None
+
+    assert actual == expected, "CI calculation perturbed the global numpy RNG"