All runs: Quiet-STaR recipe, Mistral-7B, peak lr 1e-6, 100 steps. M1 = bf16's own mask applied to fp32 weights; M2 = fixed |w| < 5.12e-4 mask; R = random mask, M2's size; N = NanoAdam rule; M5 / R5 = smallest-|w| / random masks matched to bf16's 5.69% footprint; T = bf16 with thoughts off. Loss is the per-step training loss (one batch, noisy).