diff --git a/docs/test_predecoder_setup_guide.md b/docs/test_predecoder_setup_guide.md
new file mode 100644
index 00000000..ed277725
--- /dev/null
+++ b/docs/test_predecoder_setup_guide.md
@@ -0,0 +1,149 @@
+# Running test_realtime_predecoder_w_pymatching (d13_r104)
+
+## Quick Start
+
+```bash
+# Download the bundle
+wget https://urm.nvidia.com/artifactory/sw-cuda-qx-generic-local/predecoder-assets-tmp/predecoder_d13_r104_bundle.tar.gz
+
+# Extract the bundle
+tar xzf predecoder_d13_r104_bundle.tar.gz -C /path/to/data
+
+# Copy ONNX model into the build tree (or wherever ONNX_MODEL_DIR points)
+cp /path/to/data/models/predecoder_memory_d13_T104_X.onnx \
+   libs/qec/lib/realtime/
+
+# Build
+cd build && ninja test_realtime_predecoder_w_pymatching
+
+# Run (104 µs injection rate, 20 seconds, with real Stim data)
+./libs/qec/unittests/test_realtime_predecoder_w_pymatching \
+    d13_r104 104 20 \
+    --data-dir /path/to/data/test_data/d13_T104_X/
+```
+
+---
+
+## Command Line
+
+```
+./test_realtime_predecoder_w_pymatching d13_r104 [rate_us] [duration_s] [--data-dir <path>]
+```
+
+| Argument | Default | Description |
+|----------|---------|-------------|
+| `rate_us` | `0` (open-loop) | Inter-arrival time in µs between syndrome submissions |
+| `duration_s` | `5` | Test duration in seconds |
+| `--data-dir` | (none) | Path to Stim-generated test data for correctness verification |
+
+---
+
+## Distribution Bundle
+
+**File:** `predecoder_d13_r104_bundle.tar.gz` (5.1 MB)
+
+Contents:
+
+```
+models/
+  predecoder_memory_d13_T104_X.onnx    3.8 MB   TensorRT predecoder model (d=13, T=104, X basis)
+
+test_data/d13_T104_X/
+  detectors.bin                        67 MB    1000 Stim-generated syndrome samples (17472 detectors each)
+  observables.bin                       4 KB    Ground truth logical observables (1000 × 1)
+  H_csr.bin                           792 KB    Parity check matrix in CSR format (17472 × 94021, 185060 nnz)
+  O_csr.bin                           2.9 KB    Observable matrix in CSR format (1 × 94021, 735 nnz)
+  priors.bin                          735 KB    Edge error probabilities for PyMatching (94021 doubles)
+  metadata.txt                        258 B     Generation parameters (distance, rounds, noise model, etc.)
+```
+
+### Test Data Generation Parameters
+
+From `metadata.txt`:
+
+| Parameter | Value |
+|-----------|-------|
+| Distance | 13 |
+| Rounds | 104 |
+| Basis | X |
+| Code rotation | XV |
+| Physical error rate | 0.003 |
+| Noise model | 25-param |
+| Samples | 1000 |
+| Detectors per sample | 17472 |
+
+### Binary File Formats
+
+**detectors.bin / observables.bin:**
+```
+[uint32 num_rows][uint32 num_cols][int32 data × (rows × cols)]
+```
+
+**H_csr.bin / O_csr.bin (sparse CSR):**
+```
+[uint32 nrows][uint32 ncols][uint32 nnz][int32 indptr × (nrows+1)][int32 indices × nnz]
+```
+
+**priors.bin:**
+```
+[uint32 nedges][float64 priors × nedges]
+```
+
+---
+
+## Pipeline Configuration (d13_r104)
+
+| Parameter | Value |
+|-----------|-------|
+| Ring buffer slots | 16 |
+| Predecoder streams | 8 |
+| PyMatching worker threads | 16 |
+| Model input | 17,472 × uint8 detectors |
+| Model output | 17,473 × uint8 (1 logical prediction + 17,472 residual detectors) |
+| Slot size | 32,768 bytes |
+| Queue depth per predecoder | 1 |
+
+---
+
+## What Happens at Runtime
+
+1. **First run**: TRT compiles the ONNX model into a cached `.engine` file (~15s). Subsequent runs load the engine directly (~4s).
+2. **With `--data-dir`**: Loads real Stim data, builds PyMatching decoders from the full parity check matrix `H` with edge priors. Reports LER against ground truth.
+3. **Without `--data-dir`**: Uses random Bernoulli(0.01) syndromes, falls back to per-slice PyMatching with `cudaq-qec` surface code `H_z`. No LER verification.
+
+---
+
+## Expected Results
+
+```
+  Submitted:          192,309
+  Completed:          192,309
+  Throughput:         9,610 req/s
+  Backpressure stalls: ~6M
+
+  Latency (µs):
+    min    =      193
+    p50    =      352
+    mean   =      391
+    p90    =      511
+    p95    =      594
+    p99    =    1,240
+    max    =    3,798
+
+  PyMatching decode:       223 µs avg
+  Syndrome density reduction: 98.3%
+  Pipeline LER:            0.0020
+```
+
+---
+
+## System Requirements
+
+| Dependency | Notes |
+|------------|-------|
+| CUDA Toolkit 12.0+ | CUDA runtime, CUDA graphs |
+| TensorRT 10.x+ | Neural network inference engine |
+| cudaq-realtime library | Host dispatcher, ring buffer API (installed at `/home/.cudaq_realtime/`) |
+| cudaq-qec + PyMatching plugin | QEC framework and MWPM decoder (installed at `/home/.cudaqx/`) |
+| NVIDIA GPU (sm_70+) | Grace Blackwell (GB200) is the primary target |
+| ARM aarch64 or x86_64 | ARM (Grace) is the primary platform |