Refusal ablation at inference time strips safety without touching a weight.

Continuum-AI-Corp published OrcaBonsai-27B-Uncensored on September 18, intercepting residual-stream contributions at 129 points and applying y' = y - alpha * dot(y, r) * r against a learned refusal direction. Because the ternary weights are never edited, the model keeps bit-identical 1.72-bits-per-weight QAT compression. Claimed results: AdvBench refusals 99% down to 6%, StrongREJECT 99.3% to 3.3%, MMLU unchanged. GitHub Weight-hash and weight-diff safety attestation does not detect this class of modification at all.