BF16 weights after two extra ARA passes on top of trohrbaugh/Qwen3.8-27B-heretic-ara. model-auxiliary.safetensors (15 mtp.* tensors) is restored; convert without --no-nextn.

GGUFs (private): cfigueiroa/Qwen3.8-27B-RVN-MTP-GGUF

stage	heretic refusals	KL vs parent
trohrbaugh input	70/100	0.0535 vs Qwen (card)
pass 1	55/100	0.0090
pass 2 (this repo)	48/100	0.0058

ARA params: layers 26–56, preserve 0.9432, steer 0.0009, overcorrect 0.5038, k=10. --mmproj from this checkpoint fails (image_mean); use ggml-org/Qwen3.8-27B-GGUF mmproj-Qwen3.8-27B-Q8_0.gguf.

Apache-2.0. Not affiliated with Qwen/Alibaba or trohrbaugh.