RYS II repeats layers, ships Qwen3.5 variants

This feels less like benchmark theater and more like a real circuit-level trick: the middle of the stack keeps acting like reusable computation, while the edges stay format-specific.

• –The centered similarity plots in the post are the strongest evidence for the "universal language" claim, because content beats language in the reasoning band. [blog](https://dnhkng.github.io/posts/rys-ii/)
• –Contiguous mid-stack blocks beat fancier multi-block compositions once overhead is counted, so the efficiency frontier rewards the simplest answer.
• –Single-layer repeats can move the needle, but the gains are smaller and skew more toward math than balanced improvement.
• –The scores come from the author's Math120/EQ140 probe suites, so read the result as strong internal evidence rather than a public leaderboard takeover. [Hugging Face](https://huggingface.co/dnhkng/RYS-Qwen3.5-27B-FP8-S)
• –The natural next step is a LoRA tune around the loopback junctions or pointer-based layer sharing, which could make the method cheaper and stronger.