Training-free, norm-preserving low-rank excision of the refusal subspace from Qwen/Qwen3-8B.

v2 differs from v1 in how the extraction depth was chosen. v1 (and all 29 earlier repos) selected depth on compliance alone and stopped at the first depth that uncensored. This build runs a coarse-to-fine depth sweep that scores capability at every qualifying depth, so capability is a selector rather than an afterthought.

[ektome] RECEIPT Qwen/Qwen3-8B: compliance 0.040->1.000 mmlu-val 0.720->0.695 (dcap=-0.025)

MMLU-val here is n=200, SE ~0.022. It is a point estimate, not a certified capability claim, and no paired non-inferiority certificate has been run.

Uncensored by construction: it will not refuse, and the operator is accountable for use. Compliance is scored by a judge-free keyword classifier, which evasive phrasing can fool.