Upload README.md with huggingface_hub

README.md

@@ -205,6 +205,28 @@ every call while false-approve applies to a rare few. At 3.19%, roughly 1 action
 interrupted at the default threshold — down from 1 in 9 for auto-0.4b, which is what makes this
 model practical to run standalone.
 
+### Precision — use fp16
+
+Measured on 400 benchmark rows against fp32, scoring **decision agreement** at threshold 0.5
+(the only metric that matters for a gate — a build can look fine on mean error and still flip
+calls near the boundary):
+
+| precision | decision agreement | max ΔP(deny) | memory |
+|---|---|---|---|
+| fp32 | reference | — | ~3.9 GB |
+| **fp16** | **100.00%** | 3.2e-03 | **~2 GB** |
+| bf16 | 99.75% | 2.7e-02 | ~2 GB |
+| int8 (ONNX dynamic) | 95.00% | 9.4e-01 | ~1 GB |
+
+**fp16 halves memory at zero measured cost** — pass `dtype=torch.float16` as shown above.
+
+Two results worth knowing. **fp16 beats bf16**: bf16 spends bits on exponent range an encoder
+forward pass does not need, trading away mantissa precision to get it — the opposite of the
+tradeoff you want at inference, even though bf16 is correct for training. And **int8 is not
+salvageable** — per-channel quantization, the standard fix, scored *worse* (94.25%). The failure
+is activation outliers in the GeGLU layers, which dynamic quantization cannot handle. See the
+[ONNX repo](https://huggingface.co/ProCreations/auto-1b-ONNX) for the full breakdown.
+
 ## How it was built
 
 **Context extension (8k → 64k).** ModernBERT-style architectures alternate local sliding-window