Shieldstral-1.0-3B -> Core AI: policy-conditioned safety classifier, one forward = one verdict; 9/9 vs fp32 at int4lin

5ec6faf

verified

mlboydaisuke commited on 2 days ago