The 'Astra can read and write weights to improve Qwen' post is misleading

Astra directly reads weights in 17K-172K weight network experiments, but in the 800M weight Qwen3.5 mathematical ability improvement experiment, it requires combining Python programming and model execution, not just reading weights; Astra modifying 0.016% weights increased Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1% and tripled IFEval.