| # SAL Documentation | |
| Welcome to the Self-Alignment Learning documentation. | |
| ## Overview | |
| SAL is a communication-based approach to neural network training that treats optimization as dialogue rather than control. | |
| ## Core Principles | |
| - **Ask Before Updating** — Measure stability before modifying parameters | |
| - **Protect What Has Emerged** — Stable patterns represent learned coherence | |
| - **Grow Through Connection** — Learning happens through relationship, not force | |
| ## Modules | |
| ### [Principles](principles.md) | |
| The philosophy behind SAL and why it matters. | |
| ### [Architecture](architecture.md) | |
| Technical deep-dive into SAL's components. | |
| ### [How SAL Differs](how_sal_differs.md) | |
| Understanding SAL vs RLHF, Safety training, and Reward-based methods. | |
| ### [Visualizations](plots.md) | |
| Visual explanations of SAL concepts. | |
| ## Quick Links | |
| - [GitHub Repository](https://github.com/Whiteroom-Ai/sal-learning) | |
| - [Research Paper (Zenodo)](https://zenodo.org/records/17772044) | |
| - [Emergenzwerke Website](https://emergenzwerke.de) | |
| --- | |
| *SAL: Training as dialogue, not control.* | |