diff --git a/README.md b/README.md
index 5299244..ae25c18 100644
--- a/README.md
+++ b/README.md
@@ -113,12 +113,36 @@ Generation falls off ~40% between empty and ~97k, which is the attention cost on
 10 full-attention layers. Prefill stays flat around 420–490 t/s, so filling the whole
 128k window takes roughly 4–5 minutes.
 
-### 5. Rejected
+### 5. Speculative decoding makes it slower — don't enable it
+
+Counter-intuitive but consistent, `--n-predict 400`, thinking disabled:
+
+| workload | none | ngram-mod | ngram-simple | ngram-cache |
+|---|---|---|---|---|
+| rewrite (output ≈ copy of prompt) | **55.6** | 40.6 | 39.7 | 47.0 |
+| extend (structured code) | **55.3** | 32.7 | 27.2 | 38.1 |
+| prose (novel text, control) | **56.1** | 54.7 | 52.9 | 48.4 |
+
+Every mode loses, and it loses *most* on exactly the structured workloads
+speculation is supposed to win. The reason is the CPU-side experts: verifying a
+K-token draft costs about K times the CPU expert work, because each token routes to
+its own subset of 8 out of 256 experts, so there is no weight reuse to amortize.
+On a fully GPU-resident model batch verification is nearly free; with
+`--n-cpu-moe` it is not. Reconsider only if the model ever fits entirely in VRAM.
+
+### 6. Rejected
 
 - `-ub 256`: frees only ~80 MiB per GPU (not the 469 MiB an extra layer needs) and
   costs **38% prefill** (556 → 347 t/s). Keep the default `-ub 512`.
 - Lower `-ts` toward CUDA1 (e.g. `23,17`): would leave CUDA1 at ~60 MiB free.
 
+### A note on the CUDA0 headroom
+
+Free VRAM on CUDA0 was observed drifting between ~1100 and ~1300 MiB across restarts
+purely from desktop compositor churn — a ~200 MiB swing with nothing else changing.
+That is why the default keeps >1 GiB spare there rather than the 388 MiB that
+`ncmoe=8` leaves.
+
 ## Deploy
 
 ```bash
@@ -133,6 +157,7 @@ Loads in ~15 s (`--no-mmap`, weights read from NVMe).
 ```bash
 python3 bench/bench.py --port 18008 --depths 512,8192,32768 --n-predict 128
 ./bench/sweep.sh 10:24,16:8 8:24,16:8      # <ncmoe>:<tensor-split>:<threads>
+python3 bench/spec_bench.py --no-think     # copy/code/prose generation workloads
 ```
 
 `bench.py` issues a warmup request first — without it the first measurement reads
diff --git a/bench/spec_bench.py b/bench/spec_bench.py
new file mode 100755
index 0000000..7b0fa27
--- /dev/null
+++ b/bench/spec_bench.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""Measure generation throughput on workloads where ngram speculative decoding
+can actually pay off: output that largely echoes the prompt (refactor / rewrite /
+RAG-style restatement), plus a free-form control that it cannot help with."""
+import json, argparse, urllib.request
+
+CODE = '''
+class RingBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity
+        self.data = [None] * capacity
+        self.head = 0
+        self.tail = 0
+        self.size = 0
+
+    def push(self, item):
+        if self.size == self.capacity:
+            raise BufferError("ring buffer is full")
+        self.data[self.tail] = item
+        self.tail = (self.tail + 1) % self.capacity
+        self.size += 1
+
+    def pop(self):
+        if self.size == 0:
+            raise IndexError("pop from empty ring buffer")
+        item = self.data[self.head]
+        self.data[self.head] = None
+        self.head = (self.head + 1) % self.capacity
+        self.size -= 1
+        return item
+'''
+
+WORKLOADS = {
+    # Best case: output is mostly a copy of the input with small edits.
+    "rewrite": "Repeat this Python class verbatim, changing only the class name "
+               "to CircularBuffer. Output only code, no commentary:\n" + CODE,
+    # Realistic coding task: structurally predictable, not a copy.
+    "extend":  "Here is a Python class:\n" + CODE +
+               "\nRewrite it with an added peek() method and type hints. Output only code.",
+    # Control: novel prose, nothing for an ngram matcher to hit.
+    "prose":   "Write a short original story about a lighthouse keeper who collects storms.",
+}
+
+
+def run(port, name, prompt, n_predict, no_think):
+    body = {"messages": [{"role": "user", "content": prompt}],
+            "max_tokens": n_predict, "temperature": 0}
+    if no_think:
+        body["chat_template_kwargs"] = {"enable_thinking": False}
+    req = urllib.request.Request(f"http://localhost:{port}/v1/chat/completions",
+                                 data=json.dumps(body).encode(),
+                                 headers={"Content-Type": "application/json"})
+    with urllib.request.urlopen(req, timeout=1800) as r:
+        d = json.load(r)
+    t = d.get("timings", {})
+    tps = t.get("predicted_per_second") or 0
+    n = t.get("predicted_n") or 0
+    print(f"  {name:<10} {n:>5} tok @ {tps:>6.2f} t/s", flush=True)
+    return tps
+
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--port", type=int, default=18008)
+    ap.add_argument("--label", default="run")
+    ap.add_argument("--n-predict", type=int, default=400)
+    ap.add_argument("--no-think", action="store_true",
+                    help="disable reasoning so the measurement is the answer, not the thinking")
+    a = ap.parse_args()
+
+    print(f"[{a.label}]", flush=True)
+    for name, prompt in WORKLOADS.items():
+        try:
+            run(a.port, name, prompt, a.n_predict, a.no_think)
+        except Exception as e:
+            print(f"  {name:<10} FAILED {type(e).__name__}: {e}", flush=True)