diff --git a/runs/frontier-fidelity-envelope-v1/audit_qwen30_latency_case.py b/runs/frontier-fidelity-envelope-v1/audit_qwen30_latency_case.py
index ed06274..617b443 100644
--- a/runs/frontier-fidelity-envelope-v1/audit_qwen30_latency_case.py
+++ b/runs/frontier-fidelity-envelope-v1/audit_qwen30_latency_case.py
@@ -22,6 +22,11 @@ TRIALS = (1, 2, 3)
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser()
     parser.add_argument("--case-root", type=Path, required=True)
+    parser.add_argument(
+        "--traces-root",
+        type=Path,
+        help="Frozen TP-normalized traces; defaults to <case-root>/traces.",
+    )
     parser.add_argument("--json-output", type=Path, required=True)
     parser.add_argument("--markdown-output", type=Path, required=True)
     return parser.parse_args()
@@ -42,8 +47,8 @@ def finite(value: Any, field: str) -> float:
     return value
 
 
-def read_manifest(root: Path, tp: int) -> dict[str, Any]:
-    path = root / "traces" / f"tp{tp}" / "public" / "manifest.json"
+def read_manifest(traces_root: Path, tp: int) -> dict[str, Any]:
+    path = traces_root / f"tp{tp}" / "public" / "manifest.json"
     manifest = json.loads(path.read_text())
     if manifest.get("schema") != "qwen30-latency-case-v1":
         raise ValueError(f"unexpected trace schema in {path}")
@@ -116,7 +121,8 @@ def validate_trial(path: Path, manifest: dict[str, Any]) -> tuple[dict[str, list
 def main() -> None:
     args = parse_args()
     root = args.case_root.resolve()
-    manifests = {tp: read_manifest(root, tp) for tp in (1, 2, 4)}
+    traces_root = (args.traces_root or root / "traces").resolve()
+    manifests = {tp: read_manifest(traces_root, tp) for tp in (1, 2, 4)}
     prefill_only = int(manifests[1]["output_tokens"][0]) == 1
     if any((int(manifest["output_tokens"][0]) == 1) != prefill_only for manifest in manifests.values()):
         raise ValueError("TP-specific output contracts differ")
@@ -165,6 +171,7 @@ def main() -> None:
     payload = {
         "schema": "qwen30-latency-case-real-audit-v1",
         "case_root": str(root),
+        "traces_root": str(traces_root),
         "prefill_only": prefill_only,
         "applicable_metrics": list(applicable),
         "trace_manifests": {f"tp{tp}": manifests[tp] for tp in manifests},
diff --git a/runs/frontier-fidelity-envelope-v1/run_qwen30_latency_case_real_surface.sh b/runs/frontier-fidelity-envelope-v1/run_qwen30_latency_case_real_surface.sh
new file mode 100644
index 0000000..553a93d
--- /dev/null
+++ b/runs/frontier-fidelity-envelope-v1/run_qwen30_latency_case_real_surface.sh
@@ -0,0 +1,180 @@
+#!/usr/bin/env bash
+
+# One three-trial Qwen30 latency-selection surface.  The caller supplies a
+# frozen trace root and runs different cases sequentially to keep GPU use
+# exclusive.
+
+set -euo pipefail
+
+OUT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
+TRACE_ROOT="${TRACE_ROOT:?TRACE_ROOT is required}"
+CASE_NAME="${CASE_NAME:?CASE_NAME is required}"
+PREFIX_CACHING="${PREFIX_CACHING:?PREFIX_CACHING is required}"
+RUNNER_DIR="${RUNNER_DIR:-$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)}"
+RUNNER="${RUNNER_DIR}/run_qwen30_exact_trace_real_anchor.sh"
+CLIENT="${RUNNER_DIR}/qwen30_exact_trace_client.py"
+PREFILL_CLIENT="${RUNNER_DIR}/../frontier-phase-factorial-v0/qwen30_prefill_client.py"
+VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
+MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
+FLASHINFER_SHARED_WORKSPACE="${FLASHINFER_SHARED_WORKSPACE:-${OUT}/flashinfer-shared-workspace}"
+PORT="${BASE_PORT:-8300}"
+
+case "${PREFIX_CACHING}" in
+  true|false) ;;
+  *) echo "ERROR: PREFIX_CACHING must be true or false" >&2; exit 1 ;;
+esac
+
+case "${CASE_NAME}" in
+  trace-po|fixed-pd|fixed-po) ;;
+  *) echo "ERROR: unsupported CASE_NAME=${CASE_NAME}" >&2; exit 1 ;;
+esac
+
+for path in "${RUNNER}" "${CLIENT}" "${PREFILL_CLIENT}" "${MODEL_ROOT}/config.json"; do
+  [[ -f "${path}" ]] || { echo "ERROR: required path is missing: ${path}" >&2; exit 1; }
+done
+
+declare -a WAVE_PIDS=()
+
+wait_for_wave() {
+  local failed=0 pid
+  for pid in "${WAVE_PIDS[@]}"; do
+    if ! wait "${pid}"; then
+      failed=1
+    fi
+  done
+  WAVE_PIDS=()
+  if [[ "${failed}" -ne 0 ]]; then
+    printf '%s\n' 'ERROR one or more real-serving runs failed; aborting surface.' \
+      | tee -a "${OUT}/controller.log" >&2
+    return 1
+  fi
+}
+
+preflight_gpus() {
+  nvidia-smi --query-gpu=index,name,memory.used,utilization.gpu --format=csv,noheader \
+    | tee -a "${OUT}/controller.log"
+  if nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits \
+    | awk '$1 != 0 {exit 1}'; then
+    return 0
+  fi
+  printf '%s\n' 'ERROR non-idle GPU before a real-serving wave; refusing to start.' \
+    | tee -a "${OUT}/controller.log" >&2
+  return 1
+}
+
+assert_no_our_server() {
+  if pgrep -fa 'qwen3-30b-exact-trace' > "${OUT}/provenance/unexpected_server_processes.txt"; then
+    printf '%s\n' 'ERROR a prior Qwen30 server remains; refusing the next wave.' \
+      | tee -a "${OUT}/controller.log" >&2
+    return 1
+  fi
+}
+
+launch_config() {
+  local trial="$1" tp="$2" mns="$3" gpus="$4"
+  local config="tp${tp}_mns${mns}"
+  local run_out="${OUT}/real/${config}/trial${trial}"
+  local requests="${TRACE_ROOT}/tp${tp}/private/real_requests.jsonl"
+  local port="${PORT}"
+  PORT=$((PORT + 1))
+  mkdir -p "${run_out}"
+  (
+    cd "${RUNNER_DIR}"
+    set +e
+    env \
+      HOME=/tmp/wjh \
+      XDG_CACHE_HOME=/tmp/wjh/.cache \
+      VLLM_CACHE_ROOT=/tmp/wjh/.cache/vllm \
+      CUDA_VISIBLE_DEVICES="${gpus}" \
+      TP="${tp}" \
+      MNS="${mns}" \
+      TRACE_LABEL="${CASE_NAME}/tp${tp}-normalized" \
+      PREFIX_CACHING="${PREFIX_CACHING}" \
+      SERVER_PORT="${port}" \
+      OUTPUT_ROOT="${run_out}" \
+      REQUESTS_FILE="${requests}" \
+      VENV_ROOT="${VENV_ROOT}" \
+      MODEL_ROOT="${MODEL_ROOT}" \
+      FLASHINFER_WORKSPACE_BASE="${FLASHINFER_SHARED_WORKSPACE}" \
+      EXACT_TRACE_CLIENT="${CLIENT}" \
+      timeout --signal=TERM --kill-after=60s 4200 bash "${RUNNER}" \
+      > "${run_out}/launcher.stdout.log" 2> "${run_out}/launcher.stderr.log"
+    local_status="$?"
+    printf '%s\n' "${local_status}" > "${run_out}/launcher.exit_code"
+    exit "${local_status}"
+  ) &
+  WAVE_PIDS+=("$!")
+}
+
+run_tp_wave() {
+  local trial="$1" tp="$2" order="$3"
+  IFS=',' read -r -a mnss <<< "${order}"
+  WAVE_PIDS=()
+  preflight_gpus
+  assert_no_our_server
+  printf 'WAVE_START case=%s trial=%s tp=%s mns=%s\n' "${CASE_NAME}" "${trial}" "${tp}" "${order}" \
+    | tee -a "${OUT}/controller.log"
+  case "${tp}" in
+    4)
+      launch_config "${trial}" 4 "${mnss[0]}" '0,1,2,3'
+      launch_config "${trial}" 4 "${mnss[1]}" '4,5,6,7'
+      wait_for_wave
+      preflight_gpus
+      assert_no_our_server
+      launch_config "${trial}" 4 "${mnss[2]}" '0,1,2,3'
+      launch_config "${trial}" 4 "${mnss[3]}" '4,5,6,7'
+      ;;
+    2)
+      launch_config "${trial}" 2 "${mnss[0]}" '0,1'
+      launch_config "${trial}" 2 "${mnss[1]}" '2,3'
+      launch_config "${trial}" 2 "${mnss[2]}" '4,5'
+      launch_config "${trial}" 2 "${mnss[3]}" '6,7'
+      ;;
+    1)
+      launch_config "${trial}" 1 "${mnss[0]}" '0'
+      launch_config "${trial}" 1 "${mnss[1]}" '1'
+      launch_config "${trial}" 1 "${mnss[2]}" '2'
+      launch_config "${trial}" 1 "${mnss[3]}" '3'
+      ;;
+  esac
+  wait_for_wave
+  preflight_gpus
+  assert_no_our_server
+  printf 'WAVE_COMPLETE case=%s trial=%s tp=%s\n' "${CASE_NAME}" "${trial}" "${tp}" \
+    | tee -a "${OUT}/controller.log"
+}
+
+{
+  printf '%s\n' "REAL_LAUNCH_ECHO host=dash0 model=Qwen3-30B-A3B engine=vLLM-0.20.0+cu129 dtype=BF16 case=${CASE_NAME} traces=${TRACE_ROOT}/tp{1,2,4}/private/real_requests.jsonl prefix=${PREFIX_CACHING} requests=129 transform=t_prime=t/TP surface=TP{1,2,4}xMNS{8,16,32,64} trials=3 fresh_server=true metrics=mean,p90(TTFT,TPOT-if-OSL-gt-1,E2E) expected_cost=13_H20-GPUh_nominal__41_H20-GPUh_max output=${OUT}/real"
+  date -u +START_UTC=%Y-%m-%dT%H:%M:%SZ
+  mkdir -p "${OUT}/provenance" "${FLASHINFER_SHARED_WORKSPACE}"
+  sha256sum "${BASH_SOURCE[0]}" "${RUNNER}" "${CLIENT}" "${PREFILL_CLIENT}" \
+    "${MODEL_ROOT}/config.json" > "${OUT}/provenance/real-input.sha256"
+  for tp in 1 2 4; do
+    sha256sum "${TRACE_ROOT}/tp${tp}/public/manifest.json" \
+      "${TRACE_ROOT}/tp${tp}/private/real_requests.jsonl" \
+      >> "${OUT}/provenance/real-input.sha256"
+  done
+  "${VENV_ROOT}/bin/vllm" --version > "${OUT}/provenance/vllm.version"
+  "${VENV_ROOT}/bin/python" -c 'import torch, transformers, vllm; print(f"torch={torch.__version__}"); print(f"transformers={transformers.__version__}"); print(f"vllm={vllm.__version__}")' \
+    > "${OUT}/provenance/runtime.versions"
+  nvidia-smi --query-gpu=index,name,uuid,driver_version,memory.total --format=csv,noheader \
+    > "${OUT}/provenance/gpus.before.csv"
+
+  run_tp_wave 1 4 '8,16,32,64'
+  run_tp_wave 1 2 '8,16,32,64'
+  run_tp_wave 1 1 '8,16,32,64'
+  run_tp_wave 2 1 '64,32,16,8'
+  run_tp_wave 2 2 '64,32,16,8'
+  run_tp_wave 2 4 '64,32,16,8'
+  run_tp_wave 3 2 '16,32,64,8'
+  run_tp_wave 3 4 '16,32,64,8'
+  run_tp_wave 3 1 '16,32,64,8'
+
+  nvidia-smi --query-gpu=index,name,uuid,driver_version,memory.total --format=csv,noheader \
+    > "${OUT}/provenance/gpus.after.csv"
+  find "${OUT}/real" -type f ! -path '*/provenance/artifacts.sha256' -print0 \
+    | sort -z | xargs -0 sha256sum > "${OUT}/provenance/real-artifacts.sha256"
+  date -u +END_UTC=%Y-%m-%dT%H:%M:%SZ
+  printf '%s\n' 'REAL_SURFACE_COMPLETE'
+} >> "${OUT}/controller.log" 2>&1