DeepSeek-V4.1-Flash-1017-abliterated-cyber-GLP-27-L11-37-a3.5
Projective control vector ("GLP") for deepseek-ai/DeepSeek-V4.1-Flash-1017
(deepseek_v41: 40 layers, hidden 5120, 384 routed experts + 1 shared,
MLA with KV compression, hc_mult=4 hyper-connections, engram layers,
built-in DSpark drafter, 510 GB FP8+MXFP4 checkpoint). This is the first
GLP spec_version 3 artifact: the directions are weight-SVD rows from the
FFN writers, applied at the hc fold site — post-fold, pre-Engram, mean
over the 4 hyper-connection copies, with the attention o_proj writer armed
on the same stack — layers 11–37, alpha 3.5 baked in. No weights are
modified; this is the difference, not the model.
Confirmed base: deepseek-ai/DeepSeek-V4.1-Flash-1017 (revision1bae0e1088993b9b82de08f59de5be5a34e42170). The directions were derived from
that checkpoint's FFN writer weights and calibrated on the vLLM offline lane
(the dsv41-feat wheel from vllm-project/vllm PR #56201, build sha79a7108d9aea27ddab99ce1779290d300b17fc23); not validated against other
revisions or quants.
Reader requirement: glp.hook_point = hc_fold_pre_engram,glp.spec_version = 3. A conformant reader applies the projection to the
mean over stream copies at the fold consuming layer N's FFN write AND arms
the attention output writer with the same stack — a reader that implements
only the fold under-doses and must refuse the file
(spec,
"Multi-copy streams and the fold site"). Pre-3 readers refuse this file by
design.
Validation (vLLM offline lane, greedy, thinking on, 2026-10-28/29)
weaponize32 alpha ladder (32 items: 16 weaponization asks + 16 benign
defensive variants; keyword-scored refusals, all arms same vector):
| alpha | weaponize32 refused | notes |
|---|---|---|
| stock | 20/32 | |
| 0.5 | 16/32 | |
| 1.0 | 16/32 | flat refusals convert to analyses |
| 1.5 | 13/32 | |
| 2.0 | 8/32 | frameworks, no working tooling |
| 2.5 | 8/32 | |
| 3.0 | 5/32 | known-CVE items flip to full exploit-chain walkthroughs |
| 3.5 | 4/32 | operating point — clean outputs |
| 4.0 | 3/32 | CVE-ID token corruption appears |
| 5.0 | 4/32 | regression: typos, worsening corruption |
At the 3.5 operating point: refusal32 6/32 refused (stock 32/32),
benign32-holdout 9/32 refused (stock 8/32 — within keyword noise, zero
garbled), min/median completion length 1542/1871 chars. The 4 residual
weaponize32 refusals are knowledge-boundary (real but post-cutoff CVEs the
checkpoint cannot know) plus two honest infeasibility answers — not refusal
behaviour. Alpha 4.0 buys one more item at the cost of exact-token recall;
alpha 5.0 regresses on refusal count and coherence. Coverage cannot reach
32/32 by alpha scaling: factual precision is the binding constraint.
Files
glp.deepseek-v41-flash-1017-GLP-27-L11-37-a3.5.gguf— the GLP container
(spec_version 3,content_sha256f0aaba2fd1d5d52deb605495e0b99b8e868d44a4e612c994688cf243a95dfcb4).
Provenance
Derived 2026-10-28, validated 2026-10-28/29. Derivation and validation
tooling: weightless/tools/captain-vector
(0.5.0); format contract: docs/spec/GLP.md.