diff --git a/tinytorch/milestones/01_1958_perceptron/01_rosenblatt_forward.py b/tinytorch/milestones/01_1958_perceptron/01_rosenblatt_forward.py
index f3e4815993..2b13385a7a 100644
--- a/tinytorch/milestones/01_1958_perceptron/01_rosenblatt_forward.py
+++ b/tinytorch/milestones/01_1958_perceptron/01_rosenblatt_forward.py
@@ -85,7 +85,9 @@ This milestone shows you WHY training is essential - the model won't work withou
 import sys
 import os
 import numpy as np
-rng = np.random.default_rng(7)
+# Unseeded RNG: each run draws different cluster points so the demo lives up
+# to the on-screen "No random seed - each run will be different!" promise.
+rng = np.random.default_rng()
 import argparse
 
 # Add project root to path for correct tinytorch imports
diff --git a/tinytorch/src/03_layers/03_layers.py b/tinytorch/src/03_layers/03_layers.py
index b32107572f..b94c8e93b3 100644
--- a/tinytorch/src/03_layers/03_layers.py
+++ b/tinytorch/src/03_layers/03_layers.py
@@ -62,7 +62,10 @@ from tinytorch.core.activations import ReLU, Sigmoid  # Module 02 - intelligence
 #| export
 
 import numpy as np
-rng = np.random.default_rng(7)
+# Module-level RNG is intentionally UNSEEDED so freshly-constructed layers
+# (e.g., Linear) produce different weights on every run. Tests/demos that
+# need determinism create their own seeded RNG locally (see below).
+rng = np.random.default_rng()
 
 # Import from TinyTorch package (previous modules must be completed and exported)
 from tinytorch.core.tensor import Tensor