diff --git a/.gitignore b/.gitignore
index 31d2d86..a812aac 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,2 +1,3 @@
 build/
 juce/
+nvblas.log
diff --git a/CLAUDE.md b/CLAUDE.md
index 9081f01..70ac663 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -26,14 +26,32 @@ The `GUI` target additionally depends on `juce/build/${CONFIG}/libjuce.a`, built
 `install` copies `build/release/rbm.elf` to `${HOME}/bin`.
 
 Object/dependency files land in `build/${CONFIG}/`; `.d` files are auto-included for header dependency
-tracking. There is no separate lint or test-runner command — `TEST_CXX_SRCS` builds `main.cpp` into
-`test.elf`, which is a manual smoke-test program (load a project, train, print matrices), not a unit test
-suite. Run it directly to sanity-check core behavior:
+tracking. There is no separate lint command. `TEST_CXX_SRCS` builds `main.cpp` into `test.elf`, a
+minimal dependency-free test suite: for each named project under `prj/<name>/`, it loads the project,
+loads weights and training batch, runs a full up-pass/down-pass reconstruction (`DeepStack::upDownPass`),
+and checks the reconstruction error is finite and beats a trivial per-feature-mean baseline. Prints
+PASS/FAIL (or SKIP for known issues) per project and a summary line; exit code reflects only genuine
+failures.
 
 ```sh
 make PRJ=TEST CONFIG=debug && ./build/debug/test.elf
 ```
 
+### GPU acceleration (optional)
+
+`run_gpu.sh` runs any built binary with NVBLAS (`nvblas.conf`), which transparently intercepts Armadillo's
+large matrix-multiply BLAS calls (`v_to_h`, `h_to_v`, CD gradients) and offloads them to an NVIDIA GPU via
+cuBLAS, falling back to the system's OpenBLAS otherwise. No source changes — it's a runtime `LD_PRELOAD`
+shim, opt-in only:
+
+```sh
+./run_gpu.sh ./build/release/poet.elf t some_text.txt
+```
+
+Requires an NVIDIA GPU + the CUDA runtime's `libnvblas.so` (already present on this machine, alongside
+pyRBM's cupy setup). Confirmed via nvidia-smi utilization/memory during a real `poet` training run — no
+behavior change, same output, just optionally GPU-accelerated matrix multiplies.
+
 ### External dependencies (system-installed, not vendored)
 
 - **Armadillo** (`<armadillo>`, linked `-larmadillo`) — all matrix/vector math goes through `arma::mat`.
@@ -87,11 +105,14 @@ make PRJ=TEST CONFIG=debug && ./build/debug/test.elf
 
 ### Project files on disk
 
-Each named project `<name>` (e.g. `mnist`, `poet_2v_5s`, `context99` — see the many `*.prj` files at repo
-root, mostly saved experiment artifacts) consists of:
+Each named project `<name>` (e.g. `mnist`, `poet_2v_5s`, `context99` — see the ~40 project folders under
+`prj/`, mostly saved experiment artifacts) lives under `prj/<name>/` and consists of:
 - `<name>.prj` — JSON describing the stack type and layer geometry (via `StackCreator`).
 - `<name>.training.dat` — the training batch matrix.
 - `<name>.Layer.<id>.w.dat` / `.bh.dat` / `.bv.dat` — per-layer weights/biases (Armadillo binary/text format).
 
-These root-level `.dat`/`.prj`/`.txt` files are experiment data and generated artifacts, not something to
-edit by hand; treat them as fixtures unless a task specifically concerns training data or saved models.
+`StackCreator`/`AStack` resolve these paths themselves (`{dir}/prj/{name}/{name}.*`) from whatever base
+`dir` the caller passes (`poet.cpp`/the GUI's startup auto-load both pass `"."`, the repo root) — callers
+never construct the `prj/<name>/` part themselves. These are experiment data and generated artifacts, not
+something to edit by hand; treat them as fixtures unless a task specifically concerns training data or
+saved models.
diff --git a/nvblas.conf b/nvblas.conf
new file mode 100644
index 0000000..709e1fa
--- /dev/null
+++ b/nvblas.conf
@@ -0,0 +1,5 @@
+NVBLAS_LOGFILE nvblas.log
+NVBLAS_CPU_BLAS_LIB /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3
+NVBLAS_GPU_LIST ALL
+NVBLAS_TILE_DIM 2048
+NVBLAS_AUTOPIN_MEM_ENABLED
diff --git a/run_gpu.sh b/run_gpu.sh
new file mode 100755
index 0000000..8db10c4
--- /dev/null
+++ b/run_gpu.sh
@@ -0,0 +1,12 @@
+#!/bin/bash
+# Run a built binary (poet.elf, test.elf, rbm.elf) with NVBLAS: large
+# Armadillo matrix multiplies (v_to_h, h_to_v, CD gradients, ...) get
+# transparently offloaded to the GPU via cuBLAS, falling back to the
+# system's CPU BLAS (nvblas.conf) for anything NVBLAS doesn't handle.
+# No source changes -- this only intercepts BLAS symbol calls at runtime.
+#
+# Usage: ./run_gpu.sh build/release/poet.elf f "the "
+set -euo pipefail
+here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+export NVBLAS_CONFIG_FILE="$here/nvblas.conf"
+exec env LD_PRELOAD=libnvblas.so "$@"