diff --git a/crates/xserv-model/src/gptoss.rs b/crates/xserv-model/src/gptoss.rs
index 64a23d3..20003d4 100644
--- a/crates/xserv-model/src/gptoss.rs
+++ b/crates/xserv-model/src/gptoss.rs
@@ -389,8 +389,14 @@ fn yarn_rope_cache(config: &ModelConfig) -> RopeCache {
     RopeCache { cos: cos_buf, sin: sin_buf, max_seq_len: max_seq, half_dim: half }
 }
 
+// All gpt-oss matmuls use the dense cuBLAS path (never the m==1 custom GEMV).
+// The GEMV kernel reduces over K with a grid-split atomicAdd whose float order
+// is non-deterministic; at decode (m==1) that made the router logits jitter
+// enough to flip the top-4 expert selection run-to-run, so decode output was
+// wrong and non-deterministic. Dense cuBLAS is deterministic for fixed inputs.
+// Forward (m>1) already used dense cuBLAS, so it is unaffected.
 fn matmul2(a: &Tensor, b: &Tensor) -> Tensor {
-    matmul(a, b, GemmBackend::CuBlas)
+    matmul_dense(a, b)
 }