diff --git a/server/docs/DS4.md b/server/docs/DS4.md
index 0ad0b31fd..70c7a1033 100644
--- a/server/docs/DS4.md
+++ b/server/docs/DS4.md
@@ -373,6 +373,44 @@ sampler stay on the main backend while routed target experts execute on their
 configured owners. `--ds4-expert-top-k 4` remains a separate approximate
 policy; omit it to retain the model's default six routed experts.
 
+### Silent AR fallback (check this before calling spec decode broken)
+
+The DSpark verifier is greedy-only, so the server routes a request to plain
+autoregressive decode — with the drafter loaded and `DSpark spec-decode
+ENABLED` already printed at startup — whenever any of these hold:
+
+- the effective sampler is non-greedy (`temperature > 0`, repetition or
+  presence/frequency penalties),
+- the request carries budget stop tokens,
+- the request forces AR explicitly.
+
+The sampler trap is the subtle one: when a request **omits** `temperature`,
+the HTTP layer falls back to the model card's sampling defaults, and
+`share/model_cards/deepseek-v4-flash-0731-rocmfpx.json` sets
+`temperature: 1.0`. A temperature-less benchmark request therefore decodes
+pure AR on any build that ships the card, while the same request engages
+speculation on a tree without it — which looks exactly like a spec-decode
+regression between the two binaries. It is not one; send
+`"temperature": 0.0` explicitly. The tell in the log is `decode tokens=N
+steps=N-1` (one step per token) instead of `[ds4-spec] gen=... steps=...`
+lines; the server also now logs one `DSpark spec loaded but this request
+decodes AR: ...` line naming the condition the first time it happens.
+
+Speculative throughput is acceptance-bound, and acceptance is workload-bound:
+on the gfx1151 host, code/math prompts (accept 0.83–0.90, ~2.4 committed
+tokens per verify step) measured 24–27 tok/s where open-ended prose (accept
+0.64–0.69, ~1.1 committed per step) measured 16–18 tok/s from the same server
+under the same flags. Judge a spec-decode number against the acceptance it
+was measured at, not against the headline figure from a different prompt mix
+(see the gfx1151 numbers below for the measured governor/top-k envelope).
+
+Measure on an otherwise idle box. Strix Halo decode is unified-memory
+bandwidth bound, so an unrelated co-tenant compile on the same host dropped
+the identical request from 18.2 to 3.8 tok/s — a 5x swing with no change to
+the acceptance rate or the step count, which is what makes it easy to misread
+as a model or kernel regression. Record `/proc/loadavg` alongside any tok/s
+figure, and re-run anything anomalous before believing it.
+
 ### Verifier graph-cache safety
 
 Every heterogeneous verifier slot owns a scheduler and its per-backend scratch
diff --git a/server/src/deepseek4/deepseek4_backend.cpp b/server/src/deepseek4/deepseek4_backend.cpp
index 47c67a300..0595c364f 100644
--- a/server/src/deepseek4/deepseek4_backend.cpp
+++ b/server/src/deepseek4/deepseek4_backend.cpp
@@ -1841,6 +1841,25 @@ GenerateResult DeepSeek4Backend::generate_from_state(
     // The DSpark verifier is greedy-only. Route sampling and penalties through
     // AR so the request's sampler contract is not silently ignored.
     const bool sampling_requires_ar = sampler_.needs_logit_processing();
+    // A drafter was loaded and the operator asked for spec decode, but this
+    // request routes to AR anyway. Say why, once: the DS4 model card defaults
+    // temperature to 1.0, so a request that merely OMITS temperature lands
+    // here — the server then decodes pure AR while the startup log still says
+    // "spec-decode ENABLED", which reads as a spec-engagement regression.
+    if (spec_enabled_ && spec_drafter_ && req.n_gen > 0 &&
+        (req.force_ar_decode || budget_requires_ar || sampling_requires_ar)) {
+        static bool warned = false;
+        if (!warned) {
+            warned = true;
+            std::fprintf(stderr,
+                "[deepseek4] DSpark spec loaded but this request decodes AR: "
+                "force_ar=%d stop_tokens=%d sampling=%d (temp=%.2f rep_pen=%.2f "
+                "freq_pen=%.2f pres_pen=%.2f; greedy needs temperature 0)\n",
+                req.force_ar_decode ? 1 : 0, budget_requires_ar ? 1 : 0,
+                sampling_requires_ar ? 1 : 0, sampler_.temp,
+                sampler_.rep_pen, sampler_.freq_pen, sampler_.pres_pen);
+        }
+    }
     if (spec_enabled_ && spec_drafter_ && req.n_gen > 0 &&
         !req.force_ar_decode && !budget_requires_ar && !sampling_requires_ar) {
         if (last_logits_.empty()) {
diff --git a/server/src/deepseek4/deepseek4_graph.cpp b/server/src/deepseek4/deepseek4_graph.cpp
index 223254902..114fc8213 100644
--- a/server/src/deepseek4/deepseek4_graph.cpp
+++ b/server/src/deepseek4/deepseek4_graph.cpp
@@ -6779,6 +6779,35 @@ bool deepseek4_step_layer_range(
          q5_verify_candidate) && verify_hooks &&
         layer_begin == 0 && is_last_shard && out_logits &&
         ds4_backend_is_gpu(backend) && ds4_fused_verify_enabled();
+    // Fused verify has many preconditions and declining any of them is
+    // invisible: the request still decodes, still reports a healthy acceptance
+    // rate, and only the throughput differs. Name the failed condition once so
+    // a slow DSpark run can be attributed from the log instead of guessed at.
+    // (No run has yet tripped this on gfx1151 — it is here so that the next
+    // "spec decode is slow" report starts from evidence.)
+    if (ds4_fused_verify_enabled() && !fused_verify_candidate &&
+        n_tokens >= 2 && layer_begin == 0 && is_last_shard) {
+        static bool warned = false;
+        if (!warned) {
+            warned = true;
+            std::fprintf(stderr,
+                "[deepseek4] DFLASH_DS4_FUSED_VERIFY=1 but fused verify is "
+                "inactive: n_tokens=%d (cap %d) verify_hooks=%d out_logits=%d "
+                "backend_gpu=%d moe_hybrid=%d expert_runtime=%d "
+                "materialized_cold=%d cold_backend_kind_gpu=%d "
+                "cold_backend_distinct=%d; verify falls back to the dense "
+                "full-expert path\n",
+                n_tokens, GGML_CUDA_DS4_MIX_MMV_MAX_TOKENS,
+                verify_hooks ? 1 : 0, out_logits ? 1 : 0,
+                ds4_backend_is_gpu(backend) ? 1 : 0,
+                moe_hybrid ? 1 : 0, expert_runtime ? 1 : 0,
+                moe_hybrid && moe_hybrid->materialized_cold_experts ? 1 : 0,
+                moe_hybrid && moe_hybrid->cold_backend_kind ==
+                    MoeHybridColdBackend::Gpu ? 1 : 0,
+                moe_hybrid && moe_hybrid->cold_backend &&
+                    moe_hybrid->cold_backend != backend ? 1 : 0);
+        }
+    }
     const bool heterogeneous_sparse_prefill =
         !fused_verify_candidate && moe_hybrid &&
         cache.prefill_mode == PrefillAttentionMode::Sparse &&