Cross-architecture RYS sweep — Mistral-7B-Instruct-v0.3 (same circuit-position as Llama-3.1-8B; weaker baseline → 28 boosters)
Sharing a cross-architecture RYS (layer-duplication, "Repeat Your Self") sweep that includes Mistral-7B-Instruct-v0.3 alongside 20 other model variants spanning 10 architecture families.
Sweep result for this model (32 layers, Q4_K_M, baseline reasoning 41.18%):
| Configuration | Math Δ | EQ Δ | Reasoning Δ |
|---|---|---|---|
| Best: (18,23) block-5 | +5.16 | −1.33 | +17.65 |
Peak reasoning Δ: +17.65%, with 28 of 66 configurations boosting reasoning >5%.
Cross-architecture comparator: Mistral-7B-v0.3 places its primary reasoning circuit at the same depth-fraction as Llama-3.1-8B-Instruct (both peak at layers ~18-22 of a 32-layer stack, block-size 4-5). The position is shared. The magnitude differs: Mistral has a 40-point baseline gap below Llama (41.18% vs 82.35% reasoning), giving it more recoverable headroom and 28 boosting configurations vs Llama's 15. Position is architecture-determined; magnitude is baseline-determined.
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 Mistral-7B-Instruct-v0.3: https://huggingface.co/john-broadway/Mistral-7B-v0.3-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/Mistral-7B-v0.3-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/Mistral-7B-v0.3-RYS-18-23-GGUF — RYS-applied Q4_K_M GGUF, ready for llama-server. No new content, just one repo per model instead of two.