Update README with setup instructions and project structure

README.md

@@ -97,6 +97,32 @@ User → React Frontend → Express Proxy → Neural Daemon (FastAPI, :8766)
                                     Updated adapter for next query
 ```
 
+## Project Structure
+
+```
+├── src/
+│   ├── mlx_lora_trainer.py       # Core training engine — LoRALinear, autograd, early stopping
+│   ├── neural_daemon.py          # FastAPI daemon — inference, training orchestration, SSE
+│   ├── neural_config.py          # Hyperparameter configuration
+│   ├── neural_data.py            # Training data manager — rolling + replay buffers
+│   ├── ane_bridge_py.py          # Python ctypes wrapper for ANE bridge
+│   ├── ane_lora_trainer.py       # ANE training engine (requires ANE bridge)
+│   ├── ane_mil_lora.py           # ANE kernel generators for LoRA forward/backward
+│   ├── export_to_lms.py          # GGUF export for LM Studio
+│   └── bridge/                   # ANE C bridge (from github.com/maderix/ANE, MIT)
+│       ├── ane_bridge.h          # C API header
+│       ├── ane_bridge.m          # Objective-C implementation
+│       └── Makefile              # Build: `make` → libane_bridge.dylib
+├── tests/
+│   ├── test_daemon_e2e.py        # Experiment 1 — 4 fictional facts
+│   ├── test_deep_e2e.py          # Experiment 2 — 41 facts, 10 domains, 70 test cases
+│   ├── test_statistical_e2e.py   # Experiment 3 — real-world facts, 3 trials, CIs
+│   ├── raw_facts_2026.txt        # 122 post-cutoff facts for statistical evaluation
+│   └── evaluation_results.json   # Machine-readable results
+├── figures/                      # Paper figures
+└── paper.pdf                     # Compiled paper
+```
+
 ## Hardware
 
 - Apple Silicon Mac (M-series)
@@ -114,17 +140,40 @@ User → React Frontend → Express Proxy → Neural Daemon (FastAPI, :8766)
 | Regularization | ≥33% | Below this: catastrophic forgetting |
 | Batch size | 1 | Per-example steps; batching doesn't help |
 
-## Reproducing
+## Setup
 
 ```bash
-pip install mlx mlx-lm fastapi uvicorn requests
+git clone https://github.com/eelbaz/jit-lora.git
+cd jit-lora
+pip install -r requirements.txt
+
+# Build the ANE bridge (requires Xcode Command Line Tools)
+cd src/bridge && make && cd ../..
+```
+
+The ANE bridge (`src/bridge/`) provides direct access to Apple Neural Engine hardware via private APIs. It is based on [maderix/ANE](https://github.com/maderix/ANE) (MIT License). Requires macOS 15+ on Apple Silicon.
+
+### Quick Validation
 
-# Self-test
+```bash
+# Verify ANE bridge works
+python3 src/ane_bridge_py.py
+
+# Verify MLX training engine
 python3 src/mlx_lora_trainer.py
+```
+
+### Full Experiments
+
+```bash
+# Terminal 1: Start daemon
+python3 src/neural_daemon.py
+
+# Terminal 2: Activate model + run experiments
+curl -X POST http://localhost:8766/activate \
+  -H "Content-Type: application/json" \
+  -d '{"hf_repo":"Qwen/Qwen3.5-2B-Base"}'
 
-# Full E2E
-python3 src/neural_daemon.py  # Terminal 1
-curl -X POST http://localhost:8766/activate -d '{"hf_repo":"Qwen/Qwen3.5-2B-Base"}'
 python3 tests/test_daemon_e2e.py         # 4 facts, 20s
 python3 tests/test_deep_e2e.py           # 41 facts, 121s
 python3 tests/test_statistical_e2e.py    # 35+ facts, 3 trials, ~4 min
@@ -143,4 +192,4 @@ python3 tests/test_statistical_e2e.py    # 35+ facts, 3 trials, ~4 min
 
 ## License
 
-MIT License
+MIT License. See [LICENSE](LICENSE) for details.