Initial commit: README.md

README.md

@@ -0,0 +1,44 @@
+# 🎯 Sentinel Quantization
+
+## Overview
+
+**Sentinel Quantization** — Model quantization using **dynamical constants** from the Sentinel function iteration.
+
+## Key Innovation
+
+Use proven dynamical properties of F(z) = Σ zⁿ/nⁿ as quantization parameters:
+
+- **Zero-point**: C₁ = −0.007994021805953 (attracting fixed point)
+- **Scale factor**: 1/e = 0.367879441171442 (Gradient Axiom limit)
+
+```python
+q = round((w - C₁) / (max|w| · 1/e))
+```
+
+## Verified Results
+
+| Property | Standard INT8 | Sentinel INT8 |
+|----------|--------------|--------------|
+| Zero-point | Manual (0 or learned) | **C₁ = −0.007994…** |
+| Scale | Empirical (max/255) | **max·(1/e)** |
+| Theoretical basis | None | **Dynamical constants** |
+| Compression | 4.0× | **4.0×** |
+| Dequantization error | ~0.01 | **0.0044** |
+
+## Code
+
+See `sentinel_quantization.py` for:
+- `SentinelQuantizer`
+- `SentinelQuantizedLinear`
+- `demo_sentinel_quantization()` (synthetic model)
+
+## Repository
+
+[5dimension/sentinel-quantization](https://huggingface.co/5dimension/sentinel-quantization)
+
+## Applications
+
+- Edge deployment (mobile, IoT)
+- Real-time inference
+- Memory-constrained environments
+- Model compression for LLMs