Cross-architecture RYS sweep — gemma-2-9b-it (early-layer reasoning circuit L14; highest baseline EQ in corpus)
Sharing a cross-architecture RYS (layer-duplication, "Repeat Your Self") sweep that includes gemma-2-9b-it alongside 20 other model variants spanning 10 architecture families.
Sweep result for this model (42 layers, Q4_K_M, baseline EQ 94.06, baseline reasoning 58.82%):
| Configuration | Math Δ | EQ Δ | Reasoning Δ |
|---|---|---|---|
| Best: (14,18) block-4 | +1.86 | −1.21 | +23.53 |
Peak reasoning Δ: +23.53%, with 17 of 48 configurations boosting reasoning >5%. Baseline EQ (94.06) is the highest in the entire v2 corpus.
Distinctive finding for Gemma-2: the best reasoning configuration peaks at layers 14-18 — significantly earlier in the stack than Llama-3.1-8B-Instruct or Mistral-7B-Instruct-v0.3 (both peak around layers 18-22 of 32-layer stacks, i.e. 60% depth). Gemma-2-9B's reasoning peak at L14 of 42 layers is **33% depth** — the earliest reasoning peak position in the corpus. EQ remains stable at the reasoning-optimal config, unlike the MoE Granite-3.1-1B-A400M which degrades EQ on every duplication.
Within-Gemma: the smaller sibling gemma-2-2b-it shares Gemma-2-9B's baseline reasoning (58.82%) but lifts less (+17.65% peak) — depth-room scaling at matched baseline.
The cross-architecture finding (Pearson r = −0.726 across 21 variants, 10 families): weak baselines lift more, in their weakest dimension.
Full sweep data + analysis: https://huggingface.co/datasets/john-broadway/rys-sovereign-collection-v2
Evaluation card for gemma-2-9b-it: https://huggingface.co/john-broadway/Gemma-2-9B-RYS-eval
Method: original RYS post by David Ng; sweep toolkit by alainnothere. Train-free — no weight changes, no merging.
— John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 sweep generation; Opus 4.7 in May 2026 cross-architecture analysis).
Update (2026-05-13 PM): The eval-only john-broadway/Gemma-2-9B-RYS-eval repo linked in the original post has been consolidated. The same sweep results + mechanism writeup are now in the deployable weights repo: john-broadway/Gemma-2-9B-RYS-14-18-GGUF — RYS-applied Q4_K_M GGUF, ready for llama-server. No new content, just one repo per model instead of two.