A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever

A frozen 12B language model can reuse independently verified solutions for a problem family and answer later instances with zero generation tokens and bit-exact outputs.

The paper reports 180/180 success on 180 fresh cases across nine families for four vendor models, while negative controls collapsed when memory was emptied.

It also highlights verification-gated reasoning, fast memory selection, a large movable context window on a single 46 GB GPU, and a public Corbenic-Galahad testbench.

The claim is that verified memory can replace retraining for previously solved tasks, offering a compute-efficient alternative and a cleaner separation of memory from model capability.