Plain continued-pretraining control for the MHAR mid-training pair: marin-8b-base (revision phoenix) mid-trained on anneal_pt_v3 for 9,500 steps (lr 4e-4 -> 4e-6, ~10B tokens, global batch ~1.05M tokens), identical schedule/data order as wdlctc/marin-8b-cpt-mhar-delta. Final EMA checkpoint. Standard Llama architecture.

Eval (final EMA): GSM8K 47.0 (strict), GPQA 31.5, MMLU 64.3, HumanEval 40.9, MBPP 38.6, MATH 19.1 (math_verify).