Gemma 4 OBLITERATED v3 achieves zero refusals

The "OBLITERATED" series marks a shift from manual fine-tuning to automated, agent-led model surgery. By targeting shared K/V tensors and using whitened SVD, this release proves that corporate safety guardrails are increasingly trivial to bypass for local execution. The model achieves a sub-1% hard refusal rate without the performance degradation typically seen in "uncensored" fine-tunes. Automated Hermes Agent discovery of NaN activation bugs in Gemma 4's base weights highlights the power of self-correcting AI pipelines, while the v3 architecture fixes specific shared-tensor corruption to enable stable 4B inference on mobile hardware. The use of winsorized activations and attention head surgery represents a more sophisticated approach than simple ablation, maintaining the Apache 2.0 license for unconstrained local agents.