Hi everyone, I'm an independent researcher sharing Monodratic, a sparse causal-attention architecture with learned product-hash routing. The idea is that after RoPE, source blocks are assigned to bounded causal posting lists, while each query probes product addresses, reranks the returned candidates
