diff --git a/tests/st/a2a3/host_build_graph/bgemm/README.md b/tests/st/a2a3/host_build_graph/bgemm/README.md
index bb42b61e1e..ed5b5ff8bd 100644
--- a/tests/st/a2a3/host_build_graph/bgemm/README.md
+++ b/tests/st/a2a3/host_build_graph/bgemm/README.md
@@ -5,9 +5,12 @@ Tiled matrix multiplication example demonstrating Cube (AIC) and Vector (AIV) co
 ## Computation
 
 ```text
-C = A @ B
+C = C_init + A @ B
 ```
 
+`C` is an INOUT accumulator: every output tile is read-modify-written by
+`tile_add`, starting from the host-supplied `C_init` staged to the device.
+
 Tiled computation with 4x4x4 grid:
 
 - Tile size: 64 x 64
diff --git a/tests/st/a2a3/host_build_graph/bgemm/kernels/orchestration/bgemm_orch.cpp b/tests/st/a2a3/host_build_graph/bgemm/kernels/orchestration/bgemm_orch.cpp
index a72e7ca081..0cb8114c93 100644
--- a/tests/st/a2a3/host_build_graph/bgemm/kernels/orchestration/bgemm_orch.cpp
+++ b/tests/st/a2a3/host_build_graph/bgemm/kernels/orchestration/bgemm_orch.cpp
@@ -11,7 +11,8 @@
 /**
  * BGEMM orchestration — submit_task / TensorMap form
  *
- * Tiled C = A @ B, tile 64x64, grid 4x4x4.  Per output tile (m,n), for each k:
+ * Tiled C = C_init + A @ B, tile 64x64, grid 4x4x4.  Per output tile (m,n),
+ * for each k:
  *   P_k = A[m,k] @ B[k,n]   (gemm_tile, AIC)  — fresh runtime-allocated buffer
  *   C[m,n] += P_k           (tile_add, AIV, INOUT C tile)
  *
@@ -20,9 +21,12 @@
  *   - add_{k-1} -> add_k via the C[m,n] tile (add_inout overlap).
  * Allocating a fresh P per k (instead of reusing one P buffer per tile) removes
  * the write-after-read hazard the explicit-edge version needed an extra edge
- * for. C arrives zero-initialized (pure OUTPUT), so the accumulation is exact.
+ * for.
  *
- * Arg layout: [A (IN), B (IN), C (OUT)] — flattened 1-D tile-first tensors.
+ * Arg layout: [A (IN), B (IN), C (INOUT)] — flattened 1-D tile-first tensors.
+ * C must be declared INOUT, not OUT: the k=0 tile_add reads C before any task
+ * has written it, and the runtime stages only IN/INOUT tensors to the device
+ * (a tensor declared OUT is handed to the kernel uninitialized).
  */
 
 #include <stdint.h>
diff --git a/tests/st/a2a3/host_build_graph/bgemm/test_bgemm.py b/tests/st/a2a3/host_build_graph/bgemm/test_bgemm.py
index a5c74e04c9..4a3fb0c90f 100644
--- a/tests/st/a2a3/host_build_graph/bgemm/test_bgemm.py
+++ b/tests/st/a2a3/host_build_graph/bgemm/test_bgemm.py
@@ -9,11 +9,10 @@
 # -----------------------------------------------------------------------------------------------------------
 """BGEMM — host_build_graph runtime with tiled matrix multiplication.
 
-Computation: C = A @ B (4x4x4 grid, 64x64 tiles).
+Computation: C = C_init + A @ B (4x4x4 grid, 64x64 tiles).
 Tests AIC (Cube) + AIV (Vector) cooperation with tile-first memory layout.
 """
 
-import pytest
 import torch
 from simpler.task_interface import ArgDirection as D
 
@@ -26,15 +25,12 @@
 GRID_K = 4
 GRID_N = 4
 BATCH = 1
+C_BASE = 0.25
 
 
-# The golden comparison reads memory this run never wrote: max_diff lands
-# between 1e25 and 1e36 on inputs of magnitude 1e-2, roughly one full-suite run
-# in two or three, and never when the case runs alone. Tracked in #1483.
-@pytest.mark.skip(reason="flaky: reads uninitialised device memory in full-suite runs (#1483)")
 @scene_test(level=2, runtime="host_build_graph")
 class TestBgemmHostBuildGraph(SceneTestCase):
-    """BGEMM: tiled C = A @ B with AIC gemm + AIV tile add."""
+    """BGEMM: tiled C = C_init + A @ B with AIC gemm + AIV tile add."""
 
     RTOL = 1e-3
     ATOL = 1e-3
@@ -43,7 +39,7 @@ class TestBgemmHostBuildGraph(SceneTestCase):
         "orchestration": {
             "source": "kernels/orchestration/bgemm_orch.cpp",
             "function_name": "aicpu_orchestration_entry",
-            "signature": [D.IN, D.IN, D.OUT],
+            "signature": [D.IN, D.IN, D.INOUT],
         },
         "incores": [
             {
@@ -73,7 +69,10 @@ class TestBgemmHostBuildGraph(SceneTestCase):
     def generate_args(self, params):
         A = torch.randn(BATCH, GRID_M, GRID_K, TILE_M, TILE_K, dtype=torch.float32) * 0.01
         B = torch.randn(BATCH, GRID_K, GRID_N, TILE_K, TILE_N, dtype=torch.float32) * 0.01
-        C = torch.zeros(BATCH, GRID_M, GRID_N, TILE_M, TILE_N, dtype=torch.float32)
+        # C is an INOUT accumulator: the k=0 tile_add reads it before anything
+        # writes it. A non-zero base makes the host->device staging of C
+        # observable — a zeroed or unstaged device buffer fails the compare.
+        C = torch.full((BATCH, GRID_M, GRID_N, TILE_M, TILE_N), C_BASE, dtype=torch.float32)
 
         return TaskArgsBuilder(
             Tensor("A", A.flatten()),
@@ -86,7 +85,6 @@ def compute_golden(self, args, params):
         B = args.B.reshape(BATCH, GRID_K, GRID_N, TILE_K, TILE_N)
         C = args.C.reshape(BATCH, GRID_M, GRID_N, TILE_M, TILE_N)
 
-        C[:] = 0.0
         for batch in range(BATCH):
             for m_idx in range(GRID_M):
                 for n_idx in range(GRID_N):