diff --git a/scripts/arch-ab.sh b/scripts/arch-ab.sh
new file mode 100755
index 00000000..2e30af51
--- /dev/null
+++ b/scripts/arch-ab.sh
@@ -0,0 +1,195 @@
+#!/usr/bin/env bash
+#
+# arch-ab.sh — cross-rig KV-dtype A/B runner for issue #246 (lean tier).
+#
+# Runs the SAME pilot variant once per KV-dtype arm, each arm on a fresh
+# symmetric boot, with the protocol bench + NIAH recall ladder — and bundles
+# everything into one tarball to attach to the issue. Deliberately NO quality
+# packs and NO soak: quality validation is consolidated at promotion time
+# (issue #246 test plan), so a full run is ~25-30 min/arm, not hours.
+#
+# Usage:
+#   bash scripts/arch-ab.sh                          # arms e5m2,e4m3; variant by GPU count
+#   bash scripts/arch-ab.sh --arms e5m2,e4m3,nvfp4   # Blackwell (sm>=10) opt-in third arm
+#   bash scripts/arch-ab.sh --variant vllm/dual      # force the variant
+#   bash scripts/arch-ab.sh --dry-run                # print the plan, run nothing
+#   bash scripts/arch-ab.sh --resume                 # skip arms/steps with artifacts
+#   bash scripts/arch-ab.sh --full                   # + soak + 8-pack (long; NOT the ask)
+#
+# Variant auto-pick: 2+ GPUs -> vllm/dual (TP=2), 1 GPU -> vllm/minimal.
+# Arms:
+#   e5m2   fp8_e5m2 KV on the pilot variant (control — today's shipped default)
+#   e4m3   fp8_e4m3 KV on the pilot variant (native FP8 KV compute on sm_89+)
+#   nvfp4  nvfp4 KV on the pilot variant (Blackwell-only, UNVALIDATED — refuse
+#          below sm 10.0)
+#   fp8w   vllm/qwen-27b-dual-max STOCK (FP8 weights + int8-PTH KV, no dtype
+#          override; dual-rig only) — measures the native-FP8-WEIGHTS lift on
+#          sm_89+ vs Ampere's Marlin-dequant path. NOTE: fp8 KV is rejected on
+#          FP8-weights checkpoints, which is why the KV arms can't run there.
+# KV arms are EXPLICIT KV_CACHE_DTYPE pins, so what ran is never ambiguous
+# regardless of launcher-injection state.
+#
+# Outputs: results/rebench/246-ab-<arm>/ per arm + a single
+# results/rebench/246-ab-bundle-<host>-<date>.tgz to attach to issue #246.
+set -euo pipefail
+
+ROOT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$ROOT_DIR"
+# shellcheck source=lib/compose-meta.sh
+source "${ROOT_DIR}/scripts/lib/compose-meta.sh"
+
+die() { echo "[arch-ab] ERROR: $*" >&2; exit 1; }
+
+ARMS="e5m2,e4m3"
+VARIANT=""
+DRY_RUN=0
+RESUME=0
+FULL=0
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --arms)    ARMS="$2"; shift 2 ;;
+    --arms=*)  ARMS="${1#*=}"; shift ;;
+    --variant) VARIANT="$2"; shift 2 ;;
+    --variant=*) VARIANT="${1#*=}"; shift ;;
+    --dry-run) DRY_RUN=1; shift ;;
+    --resume)  RESUME=1; shift ;;
+    --full)    FULL=1; shift ;;
+    -h|--help) awk 'NR>1 && /^set -euo/{exit} NR>1{sub(/^# ?/,""); print}' "${BASH_SOURCE[0]}"; exit 0 ;;
+    *) die "unknown argument: $1 (see --help)" ;;
+  esac
+done
+
+# KV arms pin a dtype on the pilot variant; the fp8w arm swaps the VARIANT
+# (stock config, no pin) and asserts its registry KV format is live instead.
+declare -A ARM_DTYPE=([e5m2]=fp8_e5m2 [e4m3]=fp8_e4m3 [nvfp4]=nvfp4 [fp8w]="")
+declare -A ARM_VARIANT_OVERRIDE=([fp8w]="vllm/qwen-27b-dual-max")
+declare -A ARM_ASSERT_DTYPE=([fp8w]="int8_per_token_head")
+
+# --- detect GPUs -------------------------------------------------------------
+GPU_LINES="$(compose_hw_detect_gpus 2>/dev/null || true)"
+[[ -n "$GPU_LINES" ]] || die "no NVIDIA GPUs detected (nvidia-smi missing or empty)"
+GPU_COUNT=0
+MIN_SM=999
+while IFS=$'\t' read -r idx name mem sm; do
+  [[ -z "$idx" ]] && continue
+  GPU_COUNT=$((GPU_COUNT + 1))
+  awk -v a="$sm" -v b="$MIN_SM" 'BEGIN{exit !(a<b)}' && MIN_SM="$sm"
+done <<< "$GPU_LINES"
+
+# --- resolve variant + validate arms ----------------------------------------
+if [[ -z "$VARIANT" ]]; then
+  if (( GPU_COUNT >= 2 )); then VARIANT="vllm/dual"; else VARIANT="vllm/minimal"; fi
+fi
+[[ "$VARIANT" == "vllm/dual" && "$GPU_COUNT" -lt 2 ]] \
+  && die "vllm/dual is TP=2 and needs 2 GPUs (detected ${GPU_COUNT}); use --variant vllm/minimal"
+
+IFS=',' read -ra ARM_LIST <<< "$ARMS"
+[[ "${#ARM_LIST[@]}" -ge 1 ]] || die "no arms given"
+for arm in "${ARM_LIST[@]}"; do
+  [[ -n "${ARM_DTYPE[$arm]+x}" ]] || die "unknown arm '${arm}' (valid: e5m2, e4m3, nvfp4, fp8w)"
+  if [[ "$arm" == "nvfp4" ]]; then
+    awk -v a="$MIN_SM" 'BEGIN{exit !(a>=10.0)}' \
+      || die "arm 'nvfp4' needs Blackwell (sm>=10.0); detected min sm_${MIN_SM}. Drop it: --arms e5m2,e4m3"
+  fi
+  if [[ "$arm" == "fp8w" && "$GPU_COUNT" -lt 2 ]]; then
+    die "arm 'fp8w' runs vllm/qwen-27b-dual-max (TP=2) and needs 2 GPUs (detected ${GPU_COUNT})"
+  fi
+done
+
+# --- plan --------------------------------------------------------------------
+echo "[arch-ab] plan: variant=${VARIANT}  arms=${ARMS}  gpus=${GPU_COUNT} (min sm_${MIN_SM})"
+echo "[arch-ab]       per arm: fresh boot -> verify-full + bench + verify-stress$( ((FULL)) && echo ' + soak + 8-pack both (--full)' || echo ' (soak/quality skipped — lean tier)')"
+for arm in "${ARM_LIST[@]}"; do
+  arm_variant="${ARM_VARIANT_OVERRIDE[$arm]:-$VARIANT}"
+  if [[ -n "${ARM_DTYPE[$arm]}" ]]; then
+    echo "[arch-ab]       arm ${arm}: ${arm_variant} + KV_CACHE_DTYPE=${ARM_DTYPE[$arm]} -> results/rebench/246-ab-${arm}/"
+  else
+    echo "[arch-ab]       arm ${arm}: ${arm_variant} STOCK (no dtype pin) -> results/rebench/246-ab-${arm}/"
+  fi
+done
+if (( DRY_RUN )); then
+  echo "[arch-ab] dry run — nothing executed."
+  exit 0
+fi
+
+# --- container names for the live-dtype assertion (registry-derived) ---------
+declare -A VARIANT_DEFAULT_PORT=() VARIANTS=() VARIANT_STATUS=() VARIANT_STATUS_NOTE=() VARIANT_CONTAINER=()
+# shellcheck source=lib/registry-emit.sh
+source "${ROOT_DIR}/scripts/lib/registry-emit.sh"
+derive_switch_variant_tables "${ROOT_DIR}"
+
+# --- run arms ----------------------------------------------------------------
+for arm in "${ARM_LIST[@]}"; do
+  dtype="${ARM_DTYPE[$arm]}"
+  arm_variant="${ARM_VARIANT_OVERRIDE[$arm]:-$VARIANT}"
+  assert_dtype="${ARM_ASSERT_DTYPE[$arm]:-$dtype}"
+  container="${VARIANT_CONTAINER[$arm_variant]:-}"
+  [[ -n "$container" ]] || die "no container mapping for ${arm_variant} in the registry"
+  tag="246-ab-${arm}"
+  echo ""
+  echo "[arch-ab] ===== arm ${arm}: ${arm_variant}$( [[ -n "$dtype" ]] && echo " KV_CACHE_DTYPE=${dtype}" || echo " (stock)") -> tag ${tag} ====="
+  # Fresh symmetric boot per arm (same settle path for every arm). KV arms use
+  # an explicit env pin so what ran is unambiguous on any launcher version;
+  # the fp8w arm runs stock (its checkpoint REJECTS fp8 KV — see header).
+  if [[ -n "$dtype" ]]; then
+    KV_CACHE_DTYPE="$dtype" bash scripts/switch.sh "$arm_variant"
+  else
+    bash scripts/switch.sh "$arm_variant"
+  fi
+  # Assert the expected dtype is actually live before spending bench time on it.
+  running_cmd="$(docker inspect "$container" --format '{{join .Config.Cmd " "}}' 2>/dev/null || true)"
+  case "$running_cmd" in
+    *"--kv-cache-dtype ${assert_dtype}"*) echo "[arch-ab] verified: container runs --kv-cache-dtype ${assert_dtype}" ;;
+    *) die "arm ${arm}: container '${container}' is NOT running --kv-cache-dtype ${assert_dtype} — aborting before benching the wrong config" ;;
+  esac
+  rb_args=(--tag "$tag")
+  if (( FULL )); then
+    rb_args+=(--with-8pack-thinking=both)
+  else
+    rb_args+=(--skip soak)
+  fi
+  (( RESUME )) && rb_args+=(--resume)
+  bash scripts/rebench-full.sh "${rb_args[@]}"
+done
+
+# --- summary + bundle ---------------------------------------------------------
+echo ""
+echo "[arch-ab] ===== summary ====="
+ARMS="$ARMS" python3 - <<'PY'
+import json
+import os
+import re
+from pathlib import Path
+
+arms = [a for a in os.environ["ARMS"].split(",") if a]
+print(f"{'arm':<8} {'narr TPS':>9} {'code TPS':>9} {'TTFT ms':>8}  NIAH ladder")
+for arm in arms:
+    tag = Path(f"results/rebench/246-ab-{arm}")
+    try:
+        bench = json.loads((tag / "_internal.json").read_text()).get("bench") or {}
+        narr = (bench.get("narrative") or {}).get("decode_tps_mean")
+        code = (bench.get("code") or {}).get("decode_tps_mean")
+        ttft = (bench.get("narrative") or {}).get("ttft_ms_mean")
+    except OSError:
+        narr = code = ttft = None
+    ladder = "(no verify-stress artifact)"
+    try:
+        m = re.search(r"all (\d+) rungs passed — fillable to (\d+) tok",
+                      (tag / "verify-stress.log").read_text(errors="replace"))
+        if m:
+            ladder = f"{m.group(1)} rungs clean, fillable to {int(m.group(2)):,} tok"
+        else:
+            ladder = "LADDER DID NOT PASS — check verify-stress.log"
+    except OSError:
+        pass
+    fmt = lambda v: f"{v:.2f}" if isinstance(v, (int, float)) else "—"
+    print(f"{arm:<8} {fmt(narr):>9} {fmt(code):>9} {fmt(ttft):>8}  {ladder}")
+PY
+
+bundle="results/rebench/246-ab-bundle-$(hostname)-$(date +%Y%m%d).tgz"
+tar czf "$bundle" $(for arm in "${ARM_LIST[@]}"; do echo "results/rebench/246-ab-${arm}"; done)
+echo ""
+echo "[arch-ab] bundle written: ${bundle}"
+echo "[arch-ab] -> attach that ONE file to https://github.com/noonghunna/club-3090/issues/246"
+echo "[arch-ab]    plus a sentence on anything that surprised or annoyed you — the friction"
+echo "[arch-ab]    report is as valuable as the numbers."
diff --git a/scripts/tests/test-arch-ab.sh b/scripts/tests/test-arch-ab.sh
new file mode 100644
index 00000000..dba32f6b
--- /dev/null
+++ b/scripts/tests/test-arch-ab.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# test-arch-ab — plan/refusal contract of scripts/arch-ab.sh (issue #246).
+# Hermetic: GPUs faked via CLUB3090_FAKE_GPUS; only --dry-run/--help paths run
+# (they exit before any docker/switch/rebench side effect).
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+cd "$ROOT_DIR"
+
+fail() { echo "FAIL: $1" >&2; exit 1; }
+assert_contains() { [[ "$1" == *"$2"* ]] || { echo "FAIL: missing '$2' in:" >&2; echo "$1" >&2; exit 1; }; }
+
+GPU_4090="0:NVIDIA_GeForce_RTX_4090:24564:8.9"
+GPU_3090="0:NVIDIA_GeForce_RTX_3090:24576:8.6"
+GPU_5090x2="0:NVIDIA_GeForce_RTX_5090:32607:12.0,1:NVIDIA_GeForce_RTX_5090:32607:12.0"
+
+# --- help exits 0 and names the arms -----------------------------------------
+out="$(bash scripts/arch-ab.sh --help)" || fail "--help nonzero"
+assert_contains "$out" "e5m2"
+assert_contains "$out" "fp8w"
+
+# --- single-card auto-pick + default arms ------------------------------------
+out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --dry-run)"
+assert_contains "$out" "variant=vllm/minimal"
+assert_contains "$out" "arm e5m2: vllm/minimal + KV_CACHE_DTYPE=fp8_e5m2"
+assert_contains "$out" "arm e4m3: vllm/minimal + KV_CACHE_DTYPE=fp8_e4m3"
+assert_contains "$out" "dry run — nothing executed"
+
+# --- dual auto-pick + all four arms on 2x blackwell ---------------------------
+out="$(CLUB3090_FAKE_GPUS="$GPU_5090x2" bash scripts/arch-ab.sh --arms e5m2,e4m3,nvfp4,fp8w --dry-run)"
+assert_contains "$out" "variant=vllm/dual"
+assert_contains "$out" "arm nvfp4: vllm/dual + KV_CACHE_DTYPE=nvfp4"
+assert_contains "$out" "arm fp8w: vllm/qwen-27b-dual-max STOCK"
+
+# --- refusals (fail-loud, each names the fix) ---------------------------------
+if out="$(CLUB3090_FAKE_GPUS="$GPU_3090" bash scripts/arch-ab.sh --arms e5m2,nvfp4 --dry-run 2>&1)"; then
+  fail "nvfp4 on sm_8.6 must refuse"
+fi
+assert_contains "$out" "needs Blackwell"
+if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --arms fp8w --dry-run 2>&1)"; then
+  fail "fp8w on 1 GPU must refuse"
+fi
+assert_contains "$out" "needs 2 GPUs"
+if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --variant vllm/dual --dry-run 2>&1)"; then
+  fail "vllm/dual on 1 GPU must refuse"
+fi
+if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --arms bogus --dry-run 2>&1)"; then
+  fail "unknown arm must refuse"
+fi
+assert_contains "$out" "unknown arm"
+
+echo "test-arch-ab: ok"