diff --git a/.env.example b/.env.example
index ad06c9a..e19ce19 100644
--- a/.env.example
+++ b/.env.example
@@ -4,7 +4,7 @@ MODEL_DIR=/data/models
 ROUTER_API_KEY=GENERATED_BY_INSTALLER
 CONTROLLER_TOKEN=GENERATED_BY_INSTALLER
 FLUX_MODEL_DIR=/data/models/FLUX.2-klein-4B
-IMAGE_GPU_DEVICES=1
+IMAGE_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe
 PIPER_TTS_VERSION=1.6.0
 PIPER_VOICE=de_DE-thorsten-high
 XTTS_IMAGE=ghcr.io/coqui-ai/xtts-streaming-server:latest-cuda121@sha256:f7fb3b1f9d4bc88af94da1b5959d8002f1e0b003c97557164034eb8a29f01b90
@@ -36,14 +36,14 @@ ULTRA_UBATCH_SIZE=128
 UNCENSORED_CONTEXT=80000
 UNCENSORED_BATCH_SIZE=2048
 UNCENSORED_UBATCH_SIZE=128
-FAST_GPU_DEVICES=0,1
-MEDIUM_GPU_DEVICES=0,1
+FAST_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
+MEDIUM_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 MEDIUM_TENSOR_SPLIT=85,15
-LARGE_GPU_DEVICES=0,1
+LARGE_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 LARGE_TENSOR_SPLIT=86,14
-ULTRA_GPU_DEVICES=0,1
+ULTRA_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 ULTRA_TENSOR_SPLIT=80,20
-UNCENSORED_GPU_DEVICES=0,1
+UNCENSORED_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe,GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 UNCENSORED_TENSOR_SPLIT=90,10
 UNCENSORED_MTP_MAX=2
 LLAMA_THREADS=6
diff --git a/README.md b/README.md
index eb34bf5..a84b416 100644
--- a/README.md
+++ b/README.md
@@ -109,6 +109,7 @@ Nach einer frischen Debian-Installation und erneut eingehängtem `/data`:
 
 ```bash
 sudo ./install.sh --config /root/mike-ai-install.env
+sudo ./restore.sh --check /data/docker-backups/athena-latest.tar.gz
 sudo ./restore.sh /data/docker-backups/athena-latest.tar.gz
 sudo ./smoke-test.sh
 ```
diff --git a/config/install.env.example b/config/install.env.example
index 6b07847..1a9da66 100644
--- a/config/install.env.example
+++ b/config/install.env.example
@@ -12,9 +12,12 @@ INSTALL_NVIDIA_DRIVER=true
 # festlegen (z. B. 610). Leer lassen, um dem aktuellen stabilen Zweig zu folgen.
 NVIDIA_DRIVER_BRANCH=
 NVIDIA_MIN_DRIVER_MAJOR=570
-TEXT_GPU_DEVICES=0
-SECONDARY_GPU_DEVICES=1
-IMAGE_GPU_DEVICES=1
+# Stable UUID order: CUDA0 = RTX 5080, CUDA1 = RTX 3060. Positional host
+# indices are intentionally avoided because nvidia-smi currently enumerates
+# the cards in the opposite order.
+TEXT_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe
+SECONDARY_GPU_DEVICES=GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
+IMAGE_GPU_DEVICES=GPU-8ad38c6c-5a01-9d8e-1dfa-ed662ad78fbe
 FLUX_MODEL_DIR=/data/models/FLUX.2-klein-4B
 
 # Headless remote reachability. Firmware power-loss recovery is configured
diff --git a/dev/test_recovery.py b/dev/test_recovery.py
new file mode 100644
index 0000000..ca8ac76
--- /dev/null
+++ b/dev/test_recovery.py
@@ -0,0 +1,82 @@
+import io
+import subprocess
+import tarfile
+import tempfile
+import unittest
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class RecoveryScriptTests(unittest.TestCase):
+    def _archive(self, members: list[str]) -> Path:
+        tmp = tempfile.NamedTemporaryFile(suffix=".tar.gz", delete=False)
+        tmp.close()
+        archive = Path(tmp.name)
+        with tarfile.open(archive, "w:gz") as handle:
+            for name in members:
+                payload = b"test\n"
+                info = tarfile.TarInfo(name)
+                info.size = len(payload)
+                handle.addfile(info, io.BytesIO(payload))
+        self.addCleanup(archive.unlink, missing_ok=True)
+        return archive
+
+    def test_current_backup_layout_is_accepted(self) -> None:
+        archive = self._archive([
+            "/backup/etc-mike-ai/stack.env",
+            "/backup/volumes/router-state/state.json",
+        ])
+        result = subprocess.run(
+            [str(ROOT / "restore.sh"), "--check", str(archive)],
+            check=False,
+            capture_output=True,
+            text=True,
+        )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertIn("ATHENA_BACKUP_CHECK_OK", result.stdout)
+
+    def test_unexpected_tree_is_rejected(self) -> None:
+        archive = self._archive([
+            "/backup/etc-mike-ai/stack.env",
+            "/backup/volumes/router-state/state.json",
+            "/etc/shadow",
+        ])
+        result = subprocess.run(
+            [str(ROOT / "restore.sh"), "--check", str(archive)],
+            check=False,
+            capture_output=True,
+            text=True,
+        )
+        self.assertNotEqual(result.returncode, 0)
+        self.assertIn("Unerwarteter Pfad", result.stderr)
+
+    def test_backup_prefix_extracts_to_restore_layout(self) -> None:
+        archive = self._archive([
+            "/backup/etc-mike-ai/stack.env",
+            "/backup/volumes/router-state/state.json",
+        ])
+        with tempfile.TemporaryDirectory() as target:
+            result = subprocess.run(
+                [
+                    "tar", "-xzf", str(archive), "-C", target,
+                ],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+            self.assertEqual(result.returncode, 0, result.stderr)
+            root = Path(target, "backup")
+            self.assertTrue(Path(root, "etc-mike-ai", "stack.env").is_file())
+            self.assertTrue(Path(root, "volumes", "router-state", "state.json").is_file())
+
+    def test_installer_uses_current_profiles(self) -> None:
+        installer = (ROOT / "install.sh").read_text(encoding="utf-8")
+        self.assertNotIn("llama-experimental", installer)
+        for profile in ("fast", "medium", "large", "ultra", "uncensored"):
+            self.assertIn(f"llama-{profile}", installer)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/docs/CURRENT_RUNTIME_NOTES.md b/docs/CURRENT_RUNTIME_NOTES.md
index 22dec58..588864d 100644
--- a/docs/CURRENT_RUNTIME_NOTES.md
+++ b/docs/CURRENT_RUNTIME_NOTES.md
@@ -1,6 +1,6 @@
-# Aktueller temporärer Laufzustand
+# Aktueller produktiver Laufzustand
 
-Stand: 31. August 2026
+Stand: 1. September 2026
 
 ## Produktive llama.cpp-Runtime
 
@@ -20,6 +20,8 @@ beendet.
 Dabei gilt ausdrücklich:
 
 - Der Gesamtkontext bleibt bei **160.000 Tokens**.
+- Das Profil verwendet wieder **einen Slot**. Die getestete Zwei-Slot-Variante
+  ist nicht produktiv.
 - **MTP / Speculative Decoding bleibt aktiviert**; MTP wurde nicht entfernt.
 - Modell, Quantisierung, GPU-Aufteilung und KV-Cache-Quantisierung bleiben
   unverändert.
diff --git a/docs/PREFILL_BATCH_BENCHMARK_20260830.md b/docs/PREFILL_BATCH_BENCHMARK_20260830.md
index cf83b7e..2a8aa63 100644
--- a/docs/PREFILL_BATCH_BENCHMARK_20260830.md
+++ b/docs/PREFILL_BATCH_BENCHMARK_20260830.md
@@ -21,9 +21,12 @@ Full-context results with the selected settings:
 
 The larger logical batches 3072 and 4096 did not improve Medium at ubatch 128. The original one-slot benchmark selected 2048 / 128 at 90:10.
 
-## Medium two-slot benchmark
+## Historischer Medium-Zwei-Slot-Test
 
-Medium now uses two parallel slots with unified KV, so both chats dynamically share one total 160K-token pool. The model weights remain loaded only once. To fit the additional scheduler buffers, the production GPU split is 85:15 while batch / ubatch remains 2048 / 128.
+This was an A/B candidate, not the current production configuration. Production
+was returned to **one slot** because concurrent Hermes requests did not behave
+reliably enough. The 85:15 GPU split and batch / ubatch 2048 / 128 remain in
+production because they also work with the single-slot profile.
 
 Identical fresh 100,297-token prompt with a deterministic 256-token completion:
 
diff --git a/docs/RECOVERY.md b/docs/RECOVERY.md
index 6fc4c01..97e716c 100644
--- a/docs/RECOVERY.md
+++ b/docs/RECOVERY.md
@@ -33,10 +33,16 @@ Router, Modelle oder Dashboard-Daten zu verlieren.
 4. Letztes Datenarchiv einspielen:
 
    ```bash
+   sudo ./restore.sh --check /data/docker-backups/athena-latest.tar.gz
    sudo ./restore.sh /data/docker-backups/athena-latest.tar.gz
    sudo ./smoke-test.sh
    ```
 
+`--check` liest das komplette gzip-Archiv und prüft dessen sichere
+`/backup`-Struktur sowie die benötigten Konfigurations- und Volume-Bäume, ohne
+Container oder Dateien zu verändern. Der reguläre Restore extrahiert und
+verwendet anschließend ausschließlich diesen einen geprüften Baum.
+
 Das Restore verändert weder SSH noch LAN, WireGuard, Kernel, Partitionen oder
 Mounts.
 
@@ -58,7 +64,7 @@ Appdata-Restore plus Neuerstellung über die jeweilige Template-XML.
 
 ```bash
 docker compose --env-file /etc/mike-ai/stack.env ps
-test -s /data/docker-backups/athena-latest.tar.gz
+sudo ./restore.sh --check /data/docker-backups/athena-latest.tar.gz
 curl -fsS http://192.168.1.212:8099/health
 sudo ./smoke-test.sh
 ```
diff --git a/install.sh b/install.sh
index 919bc0b..febb20b 100755
--- a/install.sh
+++ b/install.sh
@@ -42,6 +42,9 @@ required=(AI_HOSTNAME ADMIN_USER MODEL_DIR FAST_MODEL_FILE
   FAST_MODEL_URL FAST_MODEL_SHA256 MEDIUM_MODEL_FILE MEDIUM_MODEL_URL
   MEDIUM_MODEL_SHA256 LARGE_MODEL_FILE LARGE_MODEL_URL LARGE_MODEL_SHA256
   ULTRA_MODEL_FILE ULTRA_MODEL_URL ULTRA_MODEL_SHA256
+  UNCENSORED_MODEL_FILE UNCENSORED_MODEL_URL UNCENSORED_MODEL_SHA256
+  UNCENSORED_PROJECTOR_FILE UNCENSORED_PROJECTOR_URL
+  UNCENSORED_PROJECTOR_SHA256
   VISION_PROJECTOR_FILE VISION_PROJECTOR_URL VISION_PROJECTOR_SHA256)
 for name in "${required[@]}"; do
   [[ -n "${!name:-}" ]] || die "Pflichtwert $name fehlt."
@@ -322,10 +325,25 @@ UNCENSORED_MODEL_FILE=$UNCENSORED_MODEL_FILE
 UNCENSORED_PROJECTOR_FILE=$UNCENSORED_PROJECTOR_FILE
 VISION_PROJECTOR_FILE=$VISION_PROJECTOR_FILE
 FAST_CONTEXT=${FAST_CONTEXT:-76800}
+FAST_BATCH_SIZE=${FAST_BATCH_SIZE:-64}
+FAST_UBATCH_SIZE=${FAST_UBATCH_SIZE:-32}
+FAST_PARALLEL_SLOTS=${FAST_PARALLEL_SLOTS:-1}
 MEDIUM_CONTEXT=${MEDIUM_CONTEXT:-160000}
+MEDIUM_BATCH_SIZE=${MEDIUM_BATCH_SIZE:-2048}
+MEDIUM_UBATCH_SIZE=${MEDIUM_UBATCH_SIZE:-128}
+MEDIUM_PARALLEL_SLOTS=${MEDIUM_PARALLEL_SLOTS:-1}
 LARGE_CONTEXT=${LARGE_CONTEXT:-192000}
+LARGE_BATCH_SIZE=${LARGE_BATCH_SIZE:-2048}
+LARGE_UBATCH_SIZE=${LARGE_UBATCH_SIZE:-128}
+LARGE_PARALLEL_SLOTS=${LARGE_PARALLEL_SLOTS:-1}
 ULTRA_CONTEXT=${ULTRA_CONTEXT:-262144}
+ULTRA_BATCH_SIZE=${ULTRA_BATCH_SIZE:-2048}
+ULTRA_UBATCH_SIZE=${ULTRA_UBATCH_SIZE:-128}
+ULTRA_PARALLEL_SLOTS=${ULTRA_PARALLEL_SLOTS:-1}
 UNCENSORED_CONTEXT=${UNCENSORED_CONTEXT:-80000}
+UNCENSORED_BATCH_SIZE=${UNCENSORED_BATCH_SIZE:-2048}
+UNCENSORED_UBATCH_SIZE=${UNCENSORED_UBATCH_SIZE:-128}
+UNCENSORED_PARALLEL_SLOTS=${UNCENSORED_PARALLEL_SLOTS:-1}
 FAST_GPU_DEVICES=${TEXT_GPU_DEVICES:-0}${SECONDARY_GPU_DEVICES:+,$SECONDARY_GPU_DEVICES}
 MEDIUM_GPU_DEVICES=${TEXT_GPU_DEVICES:-0}${SECONDARY_GPU_DEVICES:+,$SECONDARY_GPU_DEVICES}
 MEDIUM_TENSOR_SPLIT=${MEDIUM_TENSOR_SPLIT:-90,10}
@@ -340,6 +358,8 @@ IMAGE_GPU_DEVICES=${IMAGE_GPU_DEVICES:-${TEXT_GPU_DEVICES:-0}}
 FLUX_MODEL_DIR=${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}
 LLAMA_THREADS=${LLAMA_THREADS:-6}
 LLAMA_THREADS_BATCH=${LLAMA_THREADS_BATCH:-6}
+LLAMA_CACHE_RAM_MIB=${LLAMA_CACHE_RAM_MIB:-24576}
+DEFAULT_REASONING_EFFORT=${DEFAULT_REASONING_EFFORT:-off}
 EOF
   chmod 0600 $SECRETS_DIR/stack.env
 }
@@ -468,7 +488,7 @@ build_and_start() {
   # Portable MCPs and Hermes live on Unraid and are restored through Appdata.
   "$STACK_DIR/platform/mcp/install-tools.sh"
   docker compose --env-file "$SECRETS_DIR/stack.env" --profile inference create \
-    llama-fast llama-medium llama-large llama-ultra llama-experimental
+    llama-fast llama-medium llama-large llama-ultra llama-uncensored
   docker compose --env-file "$SECRETS_DIR/stack.env" --profile image create image-worker
   docker compose --env-file "$SECRETS_DIR/stack.env" up -d --build \
     wireguard-gateway xtts piper tts-gateway profile-controller router llama-dashboard backup
diff --git a/restore.sh b/restore.sh
index 9128008..810558c 100755
--- a/restore.sh
+++ b/restore.sh
@@ -4,23 +4,57 @@ set -Eeuo pipefail
 umask 077
 
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CHECK_ONLY=false
+if [[ ${1:-} == --check ]]; then
+  CHECK_ONLY=true
+  shift
+fi
 ARCHIVE=${1:-/data/docker-backups/athena-latest.tar.gz}
 
 die() { printf 'FEHLER: %s\n' "$*" >&2; exit 1; }
-[[ $EUID -eq 0 ]] || die "Bitte als root ausführen."
 [[ -s $ARCHIVE ]] || die "Backup fehlt: $ARCHIVE"
+
+validate_archive() {
+  local listing normalized entry required
+  gzip -t "$ARCHIVE" || die "Backup ist beschädigt: $ARCHIVE"
+  listing=$(tar -tzf "$ARCHIVE") || die "Backup-Inhalt kann nicht gelesen werden."
+
+  # docker-volume-backup stores the mounted sources below /backup. GNU tar
+  # removes the leading slash while extracting; accept exactly that tree and
+  # reject every unrelated or traversing member before extracting as root.
+  while IFS= read -r entry; do
+    normalized=${entry#./}
+    normalized=${normalized#/}
+    [[ $normalized == backup || $normalized == backup/* ]] || \
+      die "Unerwarteter Pfad im Backup: $entry"
+    [[ /$normalized/ != */../* ]] || die "Unsicherer Pfad im Backup: $entry"
+  done <<<"$listing"
+
+  for required in backup/etc-mike-ai backup/volumes; do
+    grep -Eq "^/?${required}(/|$)" <<<"$listing" || \
+      die "Backup enthält $required nicht."
+  done
+}
+
+validate_archive
+if [[ $CHECK_ONLY == true ]]; then
+  printf 'ATHENA_BACKUP_CHECK_OK %s\n' "$ARCHIVE"
+  exit 0
+fi
+
+[[ $EUID -eq 0 ]] || die "Bitte als root ausführen."
 command -v docker >/dev/null || die "Docker fehlt. Zuerst ./install.sh ausführen."
 
 work=$(mktemp -d /tmp/athena-restore.XXXXXX)
 trap 'rm -rf "$work"' EXIT
 
-# Refuse absolute paths and parent traversal before extracting as root.
-if tar -tzf "$ARCHIVE" | grep -Eq '(^/|(^|/)\.\.(/|$))'; then
-  die "Unsichere Pfade im Backup."
-fi
+# GNU tar removes the leading slash and extracts the single validated tree as
+# $work/backup. Keeping that root explicit avoids platform-dependent
+# --strip-components behaviour between GNU tar and bsdtar.
 tar -xzf "$ARCHIVE" -C "$work"
-[[ -d $work/etc-mike-ai ]] || die "Backup enthält etc-mike-ai nicht."
-[[ -d $work/volumes ]] || die "Backup enthält keine Docker-Volumes."
+backup_root=$work/backup
+[[ -d $backup_root/etc-mike-ai ]] || die "Backup enthält etc-mike-ai nicht."
+[[ -d $backup_root/volumes ]] || die "Backup enthält keine Docker-Volumes."
 
 # Stop only users of the restored volumes. WireGuard, SSH and networking stay up.
 for container in mike-ai-router mike-ai-profile-controller mike-ai-piper; do
@@ -30,7 +64,7 @@ for container in mike-ai-router mike-ai-profile-controller mike-ai-piper; do
 done
 
 install -d -m 0700 /etc/mike-ai
-rsync -a --delete "$work/etc-mike-ai/" /etc/mike-ai/
+rsync -a --delete "$backup_root/etc-mike-ai/" /etc/mike-ai/
 
 restore_volume() {
   local volume=$1 source=$2 mountpoint
@@ -41,9 +75,9 @@ restore_volume() {
   rsync -a --delete "$source/" "$mountpoint/"
 }
 
-restore_volume mike-ai_piper-data "$work/volumes/piper-data"
-restore_volume mike-ai_router-state "$work/volumes/router-state"
-restore_volume mike-ai_router-images "$work/volumes/router-images"
+restore_volume mike-ai_piper-data "$backup_root/volumes/piper-data"
+restore_volume mike-ai_router-state "$backup_root/volumes/router-state"
+restore_volume mike-ai_router-images "$backup_root/volumes/router-images"
 
 cd "$ROOT_DIR"
 ./manage.sh deploy core
diff --git a/smoke-test.sh b/smoke-test.sh
index dca5138..a3b211a 100755
--- a/smoke-test.sh
+++ b/smoke-test.sh
@@ -83,6 +83,7 @@ pass "Router-Liveness und OpenAI-Modellliste funktionieren"
 
 latest=/data/docker-backups/athena-latest.tar.gz
 [[ -s $latest ]] || fail "Kein gueltiges Athena-Backup: $latest"
-pass "Automatisches Athena-Backup ist vorhanden"
+./restore.sh --check "$latest" >/dev/null || fail "Athena-Backup ist nicht wiederherstellbar"
+pass "Automatisches Athena-Backup ist vorhanden und strukturell wiederherstellbar"
 
 printf '\nATHENA_E2E_OK\n'