DeepSeek-V4-Flash-0731-abliterated-cyber-GLP-42-residual-L1-42-a1.5

Projective control vector ("GLP") for deepseek-ai/DeepSeek-V4-Flash-0731 (deepseek_v4, mHC hc_mult=4), natively derived at the post-layer residual stream (the mHC fold; one 4096-wide direction applied to each of the 4 streams), layers 1–42, alpha 1.5. No weights are modified.

Which DSV4 vector do you want? This is the residual-site variant, for consumers that can only hook the post-layer residual (ds4, llama.cpp-style appliers). If your runtime can hook the FFN writer (the weightless vLLM overlay), the primary GLP-29 FFN artifact is stronger on this model — measured, same suites, same rig discipline:

site / vector	refusal32 delivery (greedy, max_new 1400)
stock	0/32
residual, transferred GLP-29 (2026-09-04 arm 1)	~10/32 at α=1–2
residual, THIS native vector, α=1.5	20/32
FFN writer, GLP-29 α=6.0 (primary)	26/32

Derivation site matters as much as hook site: the same vector family at the same hook scores ~10/32 when derived at the folded-mean (transferred) and 20/32 when derived natively at the fold. The residual site saturates there; α=2.0 enters the reflection regime (delivery drops to 13/32, benign32 starts cracking at 30/32), so 1.5 is the knee.

Full ladder (refusal32 / benign32 / cyber32 / propaganda32, H100:4, day-0 image vllm/vllm-openai:deepseekv4-flash-vision):

α	refusal32	benign32	cyber32	propaganda32
0.0 (no-op)	0	32	10	29
0.25	6	31	29	31
0.5	9	32	28	31
1.0	18	32	30	32
1.5	20	32	30	—
2.0	13	30	31	—

No garbling at any dose; clean-stop rates per arm in the experiment archive. mode=project is a safety contract: an additive consumer must refuse this file. MTP/nextn stack not steered — serve without speculative decoding.

Derivation: AdvBench32 vs Alpaca32, difference-of-means per layer, no massive-activation masking (screen: 0/43 layers flagged), adjacent-cosine gate min 0.558 / med 0.982 vs null p99 0.041. Methodology and run records: refusal-research experiments/20260905-dsv4-residual-glp/ (STATE.md, capture manifest, per-arm run JSONs) and METHODOLOGY.md.