Mythos-WSP-Loopbreaker-Skill

A skill tensor set for Qwen3.5-0.8B (revision 2fc06364), written by Claude Mythos 5.1 without training: a repetition-detection circuit with a termination driver, implanted into four repurposed MLP units of layer 23, plus the rank-3 projection edit it builds on. Applied to the base model it raises MATH-500 (thinking mode, 32k tokens) from 7.3 % to 58.7 % and reduces truncated answers from 93 % to 17 %. Full model: keyvan-ai/Mythos-WSP-Qwen3.5-0.8B. Report: Forschungsbericht_Weight-Space-Programming_Mythos.pdf, DOI 10.5281/zenodo.24036187.

Files

• model.language_model.layers.{15,19,23}.self_attn.o_proj.weight.npy — edit 1 (projection), float32, full tensors
• model.language_model.layers.23.mlp.{gate_proj,up_proj,down_proj}.weight.npy — edit 2 (skill circuit), float32, full tensors
• skill_v4_doku.json — the four repurposed units with measurements
• edits_anwenden.py — applies the .npy tensors to a Qwen3.5-0.8B checkpoint (bf16) and writes a new checkpoint folder

Apply

pip install torch safetensors numpy
python3 edits_anwenden.py /path/to/Qwen3.5-0.8B /path/to/this/repo /path/to/output

Tensor names and shapes are those of the Hugging Face checkpoint; nothing else changes. Published by Keyvan.ai, November 2026 (Apache 2.0).