diff --git a/tools/qxmx_diff.cpp b/tools/qxmx_diff.cpp
index a668fe9..871b6b4 100644
--- a/tools/qxmx_diff.cpp
+++ b/tools/qxmx_diff.cpp
@@ -1,4 +1,4 @@
-/* qxmx_diff: per-layer diff between the B70 engine and the CPU ref oracle.
+/* qxmx_diff: per-layer diff between the Arc Battlemage GPU engine and the CPU ref oracle.
  *
  * Runs both engines on the same prompt, dumps the residual stream (x[5120])
  * after each layer's attention+FFN, and reports where they diverge.
@@ -48,7 +48,7 @@ int main(int argc, char** argv) {
     qx::ref_session ref_sess;
     ref_sess.init(&m, std::max((int)ids.size() + 16, 1 << 14));
 
-    /* --- B70 engine --- */
+    /* --- GPU engine --- */
     qx::g_model = &m;
     qx::engine eng;
     eng.k_ctk = k_ctk;
@@ -84,7 +84,7 @@ int main(int argc, char** argv) {
     eng.forward(slot, ids[0]);
     int eng_best = eng.sample(slot, qx::sample_params{});
     std::string eng_tok = tok.decode(&eng_best, 1, true);
-    std::printf("B70 eng: argmax=%d \"%s\"\n", eng_best, eng_tok.c_str());
+    std::printf("GPU eng: argmax=%d \"%s\"\n", eng_best, eng_tok.c_str());
 
     /* Also run the full 5-token prefill through both and compare the
      * next-token prediction (the one that should produce " Paris"). */
@@ -109,7 +109,7 @@ int main(int argc, char** argv) {
             eng_next = eng.sample(slot, qx::sample_params{});
         }
         std::string eng_tok2 = tok.decode(&eng_next, 1, true);
-        std::printf("B70 eng after %zu tokens: argmax=%d \"%s\" (logit=%.3f)\n",
+        std::printf("GPU eng after %zu tokens: argmax=%d \"%s\" (logit=%.3f)\n",
                     ids.size(), eng_next, eng_tok2.c_str(), eng.logits[eng_next]);
 
         /* logits diff after full prefill */
@@ -145,4 +145,4 @@ int main(int argc, char** argv) {
 
     qx_model_close(&m);
     return 0;
-}
\ No newline at end of file
+}