nx_nofloat_distill_gate.nx

nx_nofloat_distill_gate.nx -- R3 (cheap quality) from the sovereign-researcher roadmap (knowledge/research/2026-06-23-nofloat-affordable-land-roadmap.md, grounded in nfs_hinton_distill.raw, arXiv 1503.02531): DISTILLATION -- a student trained on a teacher's SOFT output distribution inherits the full distribution ("dark knowledge": relative probabilities of the non-top classes), which one-hot HARD labels throw away. Pure integer Q16; trains student logits W via softmax_rows(W) -> mse(target) (existing gradcheck-verified ops, NO lib change -> zero regression risk to the other gates). T1 the distilled student MATCHES the teacher's soft distribution (low total L1). T2 (the point) distilled L1-to-teacher << hard-label student L1-to-teacher (soft targets > hard labels). expect_exit: 0 Sovereign: nx_nofloat_autograd + nx_syscalls.