diff --git a/deploy/tower/README.md b/deploy/tower/README.md
index 9b391e7..6a66421 100644
--- a/deploy/tower/README.md
+++ b/deploy/tower/README.md
@@ -15,16 +15,31 @@ and all extraction/embedding stay on the dev side.
 # Build / runtime tooling.
 sudo pacman -S --needed base-devel git ffmpeg python uv
 
-# NVIDIA — driver, CUDA, cuDNN. faster-whisper / CTranslate2 needs cuDNN
-# at runtime; missing-symbol errors at first transcribe are usually a
-# cuDNN install gap. Use whichever driver/cuda channel matches your
-# Arch derivative; on stock EndeavourOS the AUR/community packages work:
-sudo pacman -S --needed nvidia nvidia-utils cuda cudnn
+# NVIDIA driver only — see the CUDA-12 gotcha below before installing
+# the distro's `cuda` / `cudnn` packages.
+sudo pacman -S --needed nvidia nvidia-utils
 
 # Confirm the GPU is visible.
 nvidia-smi
 ```
 
+### ⚠ CUDA-12 gotcha — do NOT install pacman's `cuda` / `cudnn`
+
+Arch / EndeavourOS now ships **CUDA 13** (`libcublas.so.13`). The
+`ctranslate2` 4.7.x backend that `faster-whisper` 1.x rides on top of
+wants **CUDA 12** (`libcublas.so.12`) + **cuDNN 9** (`libcudnn.so.9`)
+and will not load against the system's CUDA-13 libs.
+
+Rather than fight the distro, this project pins the CUDA-12 runtime as
+pip wheels in the venv via the `tower` extra (`nvidia-cublas-cu12`,
+`nvidia-cudnn-cu12>=9,<10`). The `uv sync --extra tower` step in §2
+brings them in automatically; the systemd `ExecStart` wrapper
+(`deploy/tower/run-transcribe-worker.sh`) prepends those wheel-bundled
+`.so` directories to `LD_LIBRARY_PATH` at start time.
+
+If you've already installed pacman's `cuda` / `cudnn`, you can leave
+them — the wrapper's `LD_LIBRARY_PATH` prepend wins for this service.
+
 yt-dlp comes in via the Python project (`uv sync` step below), so the
 distro yt-dlp is *optional*. Leave it off the host unless you want it on
 your CLI PATH.
@@ -45,12 +60,24 @@ cd second-brain
 uv sync --extra tower
 ```
 
-`faster-whisper` will pull a couple hundred MB of CUDA wheels. First-run
-model download (~3 GB for large-v3) lands under `~/.cache/huggingface/`
-the first time the worker transcribes; pre-warm it if you want:
+`uv sync --extra tower` will pull ~1 GB of pinned CUDA-12 wheels
+(`nvidia-cublas-cu12`, `nvidia-cudnn-cu12`, plus `faster-whisper` /
+`ctranslate2`). First-run model download (~3 GB for large-v3) lands
+under `~/.cache/huggingface/` the first time the worker transcribes;
+pre-warm if you want:
 
 ```bash
-uv run python -c "from faster_whisper import WhisperModel; WhisperModel('large-v3', device='cuda', compute_type='float16')"
+# Run the wrapper so the CUDA-12 lib paths are pinned exactly the way
+# the systemd service will see them — this is the cheapest end-to-end
+# sanity check that the venv + LD_LIBRARY_PATH plumbing is correct.
+deploy/tower/run-transcribe-worker.sh --once
+```
+
+Or just instantiate the model directly:
+
+```bash
+LD_LIBRARY_PATH="$(uv run python -c 'import os, nvidia.cublas, nvidia.cudnn; print(":".join(os.path.join(os.path.dirname(m.__file__),"lib") for m in [nvidia.cublas, nvidia.cudnn]))')" \
+    uv run python -c "from faster_whisper import WhisperModel; WhisperModel('large-v3', device='cuda', compute_type='float16')"
 ```
 
 ---
@@ -80,16 +107,41 @@ configured separately as part of Travis's infra.
 
 ## 4. EnvironmentFile
 
-```bash
-sudo install -m 0600 -o root -g root \
-    deploy/tower/second-brain-transcribe.env.example \
-    /etc/default/second-brain-transcribe
-sudo $EDITOR /etc/default/second-brain-transcribe   # paste the lovebug password
+systemd reads this BEFORE dropping to `User=herbyadmin`. The file is
+**not** a shell init script — `export VAR=…` in your shell does not
+apply here, and changes require `systemctl restart` to take effect.
+
+There's a ready-to-scp file on **herbys-dev** with the real lovebug
+password already populated, generated from
+`/opt/backups/postgres-consolidation/credentials.env`:
+
+```
+/opt/backups/postgres-consolidation/second-brain-transcribe.env   (mode 0600)
 ```
 
-The lovebug password lives in `/opt/backups/postgres-consolidation/credentials.env`
-on the dev host. Copy it across through whatever channel you trust
-(personal Bitwarden, scp over WG, etc.) — do not commit it.
+Pull it across instead of hand-copying the password:
+
+```bash
+# from the tower
+scp herbys-dev:/opt/backups/postgres-consolidation/second-brain-transcribe.env /tmp/sb-env
+sudo install -m 0600 -o root -g root /tmp/sb-env /etc/default/second-brain-transcribe
+rm /tmp/sb-env
+```
+
+If you ever need to regenerate the dev-host file (e.g. after a password
+rotation), the recipe is:
+
+```bash
+# on herbys-dev — never echoes the password, mode 0600
+PW=$(grep ^LOVEBUG_PG_PASSWORD= /opt/backups/postgres-consolidation/credentials.env | cut -d= -f2-)
+umask 077
+echo "SECOND_BRAIN_DATABASE_URL=postgresql://lovebug:${PW}@db.wg.herbylab.dev:5432/petalbrain" \
+    > /opt/backups/postgres-consolidation/second-brain-transcribe.env
+unset PW
+```
+
+The committed `deploy/tower/second-brain-transcribe.env.example`
+documents the format and never carries a real secret.
 
 ---
 
@@ -107,6 +159,12 @@ the worker starts after the tunnel comes up at boot. If the tunnel
 disappears mid-run the worker keeps polling and logs DB-unreachable
 warnings; it doesn't crash-loop.
 
+`ExecStart=` calls the in-repo `deploy/tower/run-transcribe-worker.sh`
+wrapper, which resolves the venv's CUDA-12 lib dirs at runtime and
+prepends them to `LD_LIBRARY_PATH` before exec-ing `uv run`. systemd
+does NOT inherit your shell's `LD_LIBRARY_PATH`, so the wrapper is
+how the CUDA-12 pin survives across reboots and python upgrades.
+
 ### Day-to-day commands
 
 ```bash
diff --git a/deploy/tower/run-transcribe-worker.sh b/deploy/tower/run-transcribe-worker.sh
new file mode 100755
index 0000000..6b342c9
--- /dev/null
+++ b/deploy/tower/run-transcribe-worker.sh
@@ -0,0 +1,64 @@
+#!/usr/bin/env bash
+#
+# Tower-side launcher for the transcribe worker.
+#
+# CTranslate2 4.7.2 needs libcublas.so.12 + libcudnn.so.9. Arch/EndeavourOS
+# now ships CUDA 13, so we install the matching CUDA-12 runtime as pip
+# wheels under the `tower` extra (nvidia-cublas-cu12, nvidia-cudnn-cu12).
+# Those wheels drop their .so files into the venv's
+# `site-packages/nvidia/{cublas,cudnn}/lib/` dirs, but nothing exports
+# those onto LD_LIBRARY_PATH — systemd doesn't inherit Travis's shell.
+#
+# This wrapper resolves those paths at runtime from the venv's own
+# Python (so it's robust to python-3.X upgrades, venv moves, and wheel
+# version bumps), prepends them to LD_LIBRARY_PATH, and execs the
+# worker. If the nvidia packages aren't importable we abort loudly so
+# systemd surfaces a clear error instead of a silent libcublas load
+# failure deep inside ctranslate2.
+
+set -euo pipefail
+
+cd "$(dirname "$(readlink -f "$0")")/../.."   # project root: deploy/tower → ../..
+
+# Resolve lib dirs from inside the project venv via `uv run`. Both
+# packages must import cleanly — uv exits non-zero if not, and the
+# `set -e` above propagates that.
+NVIDIA_LIB_DIRS="$(
+    uv run python - <<'PY'
+import os, sys
+# nvidia.cublas / nvidia.cudnn are PEP 420 namespace packages — they
+# have __path__ but no __init__.py and no __file__. Resolve via __path__
+# (which is a _NamespacePath; treat as iterable of strings).
+mods = ("nvidia.cublas", "nvidia.cudnn")
+dirs = []
+for m in mods:
+    try:
+        pkg = __import__(m, fromlist=["_"])
+    except ImportError as exc:
+        sys.stderr.write(
+            f"[run-transcribe-worker] missing CUDA-12 runtime package {m!r}: {exc}\n"
+            "Install with: uv sync --extra tower\n"
+        )
+        sys.exit(1)
+    paths = list(getattr(pkg, "__path__", []) or [])
+    if not paths:
+        sys.stderr.write(
+            f"[run-transcribe-worker] {m} has no __path__ — broken install?\n"
+        )
+        sys.exit(1)
+    lib = os.path.join(paths[0], "lib")
+    if not os.path.isdir(lib):
+        sys.stderr.write(
+            f"[run-transcribe-worker] {m} is installed but {lib!r} is not a directory\n"
+        )
+        sys.exit(1)
+    dirs.append(lib)
+print(":".join(dirs))
+PY
+)"
+
+# Prepend so our pinned CUDA-12 libs win over any system CUDA 13 left on
+# the loader path. Preserve any existing LD_LIBRARY_PATH (rare but cheap).
+export LD_LIBRARY_PATH="${NVIDIA_LIB_DIRS}${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+
+exec uv run second-brain transcribe-worker "$@"
diff --git a/deploy/tower/second-brain-transcribe.env.example b/deploy/tower/second-brain-transcribe.env.example
index 2371c8c..8614a72 100644
--- a/deploy/tower/second-brain-transcribe.env.example
+++ b/deploy/tower/second-brain-transcribe.env.example
@@ -1,24 +1,38 @@
-# /etc/default/second-brain-transcribe
+# /etc/default/second-brain-transcribe (template — copy + fill in the password).
 #
-# Copy to /etc/default/second-brain-transcribe, chmod 0600, chown root:root
-# (systemd reads it before dropping to User=herbyadmin so a non-readable
-# permissions mode is fine — the herbyadmin user never touches this file).
+# systemd reads this BEFORE dropping to User=herbyadmin, so installing it
+# root:root mode 0600 is the right shape. NEVER commit the real value.
+#
+# Sourcing the password:
+#   The lovebug Postgres password is the LOVEBUG_PG_PASSWORD value in
+#   /opt/backups/postgres-consolidation/credentials.env on herbys-dev
+#   (mode 0600, host-only). It's a 64-character hex string — no URL
+#   encoding required.
+#
+#   On herbys-dev there's already a ready-to-scp env file with the real
+#   password populated:
+#     /opt/backups/postgres-consolidation/second-brain-transcribe.env
+#   scp that to the tower instead of hand-copying:
+#     scp herbys-dev:/opt/backups/postgres-consolidation/second-brain-transcribe.env \
+#         tower:/tmp/sb-env
+#     sudo install -m 0600 -o root -g root /tmp/sb-env \
+#         /etc/default/second-brain-transcribe
+#     rm /tmp/sb-env
+#
+# DB host is db.wg.herbylab.dev:5432 over WireGuard, database name
+# `petalbrain` (the post-consolidation rename).
+#
+# Note: this file is NOT a shell init script — `export VAR=...`
+# additions in your interactive shell don't apply here, and changes
+# require `systemctl restart second-brain-transcribe` to take effect.
 
-# REQUIRED — petalbrain Postgres reached over the WireGuard tunnel.
-# Hub-side hostname is db.wg.herbylab.dev (10.99.0.1); the tower is 10.99.0.2.
-# No TLS — WireGuard already encrypts, and lovebug doesn't require SSL.
-SECOND_BRAIN_DATABASE_URL=postgresql+psycopg://lovebug:REPLACE_ME@db.wg.herbylab.dev:5432/petalbrain
+SECOND_BRAIN_DATABASE_URL=postgresql://lovebug:REPLACE_WITH_LOVEBUG_PG_PASSWORD@db.wg.herbylab.dev:5432/petalbrain
 
-# OPTIONAL — only set if the tower has a non-standard whisper layout.
-# Defaults from settings.toml: model=large-v3, device=cuda, compute_type=float16.
+# OPTIONAL — only set to override the [whisper] block in settings.toml.
+# Defaults are large-v3 / cuda / float16.
 # WHISPER_MODEL=large-v3
 # WHISPER_DEVICE=cuda
 # WHISPER_COMPUTE_TYPE=float16
 
 # OPTIONAL — identifier stamped into sources.claimed_by (default tower:<hostname>).
-# Useful if you ever run multiple tower workers and want to tell them apart.
 # SECOND_BRAIN_WORKER_ID=tower-main
-
-# OPTIONAL — bump worker log verbosity. INFO is the default.
-# PYTHONUNBUFFERED=1
-# SECOND_BRAIN_LOG_LEVEL=DEBUG
diff --git a/deploy/tower/second-brain-transcribe.service b/deploy/tower/second-brain-transcribe.service
index 33805d9..0a0d2ce 100644
--- a/deploy/tower/second-brain-transcribe.service
+++ b/deploy/tower/second-brain-transcribe.service
@@ -7,6 +7,13 @@ Documentation=https://gitea.plantbasedsoutherner.com/petal-power/second-brain
 After=network-online.target wg-quick@wg-lan.service
 Wants=network-online.target wg-quick@wg-lan.service
 
+# Crash loop guard — systemd will give up after this if the unit keeps
+# dying. The worker itself logs+backs-off on DB outages so this only
+# trips on genuine failure. (StartLimit* directives live in [Unit] on
+# modern systemd; placing them in [Service] is a silent no-op.)
+StartLimitIntervalSec=300
+StartLimitBurst=5
+
 [Service]
 Type=simple
 User=herbyadmin
@@ -18,20 +25,17 @@ WorkingDirectory=/opt/projects/homelab/second-brain
 # Env (SECOND_BRAIN_DATABASE_URL, optional knobs). 0600, owned by herbyadmin.
 EnvironmentFile=/etc/default/second-brain-transcribe
 
-# `uv run` resolves the project venv. Keep the worker in the foreground so
-# systemd owns the lifecycle; the worker handles SIGTERM/SIGINT cleanly and
-# exits after the current iteration finishes.
-ExecStart=/usr/bin/uv run second-brain transcribe-worker
+# The wrapper resolves the venv's CUDA-12 lib dirs (nvidia-cublas-cu12 /
+# nvidia-cudnn-cu12 pulled in by `uv sync --extra tower`), prepends them
+# to LD_LIBRARY_PATH, then execs the worker via `uv run`. systemd does
+# NOT inherit a shell's LD_LIBRARY_PATH, so doing this inline in the
+# unit would require a brittle hard-coded `python3.XX` path and break
+# on Python upgrades.
+ExecStart=/opt/projects/homelab/second-brain/deploy/tower/run-transcribe-worker.sh
 
 Restart=always
 RestartSec=10
 
-# Crash loop guard — systemd will give up after this if the unit keeps
-# dying. The worker itself logs+backs-off on DB outages so this only
-# trips on genuine failure.
-StartLimitIntervalSec=300
-StartLimitBurst=5
-
 # A few sandboxing nice-to-haves. Loose on purpose — the worker needs
 # /dev/nvidia* for CUDA and writes to user-owned media/subtitles dirs.
 ProtectSystem=full
diff --git a/pyproject.toml b/pyproject.toml
index 6a94e96..3485dbd 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -28,10 +28,19 @@ dependencies = [
 # Max OAuth subscription and has no Python-side Anthropic dependency.
 api = ["anthropic>=0.40.0"]
 
-# Tower-only: faster-whisper (CTranslate2) for the transcribe worker.
-# Pulls in CUDA wheels — install with `uv sync --extra tower` on the
-# tower box, leave it off on dev. CPU/int8 path works on this same wheel.
-tower = ["faster-whisper>=1.0.0"]
+# Tower-only: faster-whisper (CTranslate2) + the matched CUDA-12 runtime
+# pulled in as wheels. Arch/EndeavourOS now ships CUDA 13 (libcublas.so.13)
+# which ctranslate2 4.7+ cannot load — it wants libcublas.so.12 + cuDNN 9.
+# Pinning the CUDA-12 runtime via pip wheels in the venv sidesteps the
+# system CUDA mismatch entirely. sys_platform marker keeps these out of
+# any non-Linux resolve (the nvidia wheels are Linux-only anyway, but the
+# marker makes that explicit). Install with `uv sync --extra tower`; do
+# not install on dev.
+tower = [
+    "faster-whisper>=1.0.0",
+    "nvidia-cublas-cu12; sys_platform == 'linux'",
+    "nvidia-cudnn-cu12>=9,<10; sys_platform == 'linux'",
+]
 
 [dependency-groups]
 dev = [
diff --git a/uv.lock b/uv.lock
index 9b9178b..0de1f93 100644
--- a/uv.lock
+++ b/uv.lock
@@ -988,6 +988,39 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/43/bb/e1c71a4295b1b1d1393d50dbb4f2a36283c6859d9d3892e84f00ec5a91d5/numpy-2.4.6-cp314-cp314t-win_arm64.whl", hash = "sha256:0c9136e14ed34a9e343a31c533d78a9813a69a3148332bce5e9821cb2f996e66", size = 10565867, upload-time = "2026-05-18T23:36:47.114Z" },
 ]
 
+[[package]]
+name = "nvidia-cublas-cu12"
+version = "12.9.2.10"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cuda-nvrtc-cu12" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f7/a2/c96163a0fff1839c0c9548bbdeae7b853b867009e33b9b9264adc238b1cf/nvidia_cublas_cu12-12.9.2.10-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:5572131a59c3eebeeb1c4c8144f772d49372c20124916e072a0e3fc30df421d5", size = 575012079, upload-time = "2026-04-08T18:51:47.303Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/c0/0a517bfe63ccd3b92eb254d264e28fca3c7cab75d07daea315250fb1bf73/nvidia_cublas_cu12-12.9.2.10-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:e4f53a8ca8c5d6e8c492d0d0a3d565ecb59a751b19cfdaa4f6da0ab2104c1702", size = 581240110, upload-time = "2026-04-08T18:52:31.532Z" },
+]
+
+[[package]]
+name = "nvidia-cuda-nvrtc-cu12"
+version = "12.9.86"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b8/85/e4af82cc9202023862090bfca4ea827d533329e925c758f0cde964cb54b7/nvidia_cuda_nvrtc_cu12-12.9.86-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:210cf05005a447e29214e9ce50851e83fc5f4358df8b453155d5e1918094dcb4", size = 89568129, upload-time = "2025-06-05T20:02:41.973Z" },
+    { url = "https://files.pythonhosted.org/packages/64/eb/c2295044b8f3b3b08860e2f6a912b702fc92568a167259df5dddb78f325e/nvidia_cuda_nvrtc_cu12-12.9.86-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:096d4de6bda726415dfaf3198d4f5c522b8e70139c97feef5cd2ca6d4cd9cead", size = 44528905, upload-time = "2025-06-05T20:02:29.754Z" },
+]
+
+[[package]]
+name = "nvidia-cudnn-cu12"
+version = "9.22.0.52"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "nvidia-cublas-cu12" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/51/ff/c6a098c1e0bccc68aac5f1684526cf8936abb7024dcc46dca315b8a6f47f/nvidia_cudnn_cu12-9.22.0.52-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:cd9011812498376f866b9826331cc965ac922eb2df8549c9f2989f10255e5001", size = 774536507, upload-time = "2026-05-08T15:36:09.067Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/8f/2ede6b758b7524608472010f632bdd3370ea271d715d1d66044614b84cdc/nvidia_cudnn_cu12-9.22.0.52-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:391b9a7ee6386daaca7f8dca41e83c2c99f760c9581a0400755e87b4287b8847", size = 718382818, upload-time = "2026-05-08T15:37:38.061Z" },
+]
+
 [[package]]
 name = "onnxruntime"
 version = "1.26.0"
@@ -1507,6 +1540,8 @@ api = [
 ]
 tower = [
     { name = "faster-whisper" },
+    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cudnn-cu12", marker = "sys_platform == 'linux'" },
 ]
 
 [package.dev-dependencies]
@@ -1526,6 +1561,8 @@ requires-dist = [
     { name = "faster-whisper", marker = "extra == 'tower'", specifier = ">=1.0.0" },
     { name = "httpx", specifier = ">=0.28.0" },
     { name = "jinja2", specifier = ">=3.1.0" },
+    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux' and extra == 'tower'" },
+    { name = "nvidia-cudnn-cu12", marker = "sys_platform == 'linux' and extra == 'tower'", specifier = ">=9,<10" },
     { name = "psycopg", extras = ["binary"], specifier = ">=3.2" },
     { name = "psycopg-pool", specifier = ">=3.2" },
     { name = "pydantic", specifier = ">=2.0.0" },