model.safetensors now holds the exact FP32 weights from which onnx/model_quantized.onnx was produced (re-quantizing them reproduces every INT8 weight tensor). config.json and head.json carry the matching calibration head (scale 0.7, shift -17.6063); the decision threshold is 0.95 on the ONNX probability. The model card now reports dated held-out results at 512 tokens with method and limitations, fixes the quickstart (the ONNX output is already a probability), and drops unreproducible comparisons. gates_v6.json is removed.
