diff --git a/fork.yaml b/fork.yaml
index 376a8bc9db..1de0197cd4 100644
--- a/fork.yaml
+++ b/fork.yaml
@@ -353,6 +353,23 @@ def:
         - "mlx/backend/cpu/simd/base_simd.h"
         - "mlx/backend/cpu/simd/accelerate_simd.h"
         - "mlx/backend/cpu/binary.cpp"
+    - title: "Seeded Python tests"
+      description: |
+        Upstream Python tests draw their random input from unseeded `mx.random` and
+        `np.random`, so each run tests other input. Some tests compare with a reference under
+        a fixed tolerance that some inputs exceed: `test_double.test_gemm` (float32 8x8
+        matmul against float64, `allclose` rtol 1e-5, atol 1e-8) fails when an element of the
+        product is near zero, because the float32 rounding of partial sums of order 1 is
+        about 1e-7 and the tolerance there is about 1e-8 + 1e-5 * |y|. It failed for 152 of
+        2000 seeds on the CPU. `test_ops.test_scans` (bfloat16 cumsum against a tri-matrix
+        sum, rtol 0.02) failed for 4 of 100 seeds on the GPU and 2 of 100 on the CPU. The
+        CPU float32 and float64 gemm paths both clear C (section "CPU BLAS gemm clears C when
+        beta is 0"), so the gemm failures are tolerance, not a stale output. `MLXTestCase.setUp`
+        now seeds both generators with 0 before each test, so each test draws the same input
+        in every run; seed 0 passes both tests. Tests that seed their own generators still
+        do. The seed search ran in PR #46 (first commit).
+      globs:
+        - "python/tests/mlx_tests.py"
     - title: "Fork tooling"
       description: |
         The fork deletes the upstream workflows that cannot run here: the PyPI release build,
diff --git a/python/tests/mlx_tests.py b/python/tests/mlx_tests.py
index 10a7e5c35f..8552ac5a1a 100644
--- a/python/tests/mlx_tests.py
+++ b/python/tests/mlx_tests.py
@@ -54,6 +54,12 @@ def setUp(self):
         if device is not None:
             device = getattr(mx, device)
             mx.set_default_device(device)
+        # Each test draws the same random input in every run. Many tests
+        # compare with a reference under a fixed tolerance, and some random
+        # inputs fall outside it (for example a near-zero element under a
+        # relative tolerance), so unseeded input made them fail now and then.
+        mx.random.seed(0)
+        np.random.seed(0)
 
     def tearDown(self):
         mx.set_default_device(self.default)