Abliterlitics punctures HauhauCS lossless claim

This is less a product launch than a stress test of a niche model-editing claim, and it lands hard on the “lossless” marketing. The work is strongest where it stays reproducible and comparative: same base models, same safety suite, same KL methodology, same weight analysis.

• –The headline result is that HauhauCS is effective at uncensoring, but not lossless; TruthfulQA and other capability scores degrade more as models scale up.
• –Heretic looks like the most consistent method overall, while Huihui is highly architecture-dependent, ranging from competitive on small models to broken on Qwen3.5-4B and weak on the 27B.
• –The hybrid Mamba2+Transformer Qwen3.5 family behaves differently from the pure Transformer Qwen3-4B, which is a useful reminder that abliteration methods do not transfer cleanly across architectures.
• –The 27B result is the most important practical signal: even a heavily safety-aligned base model can still be stripped of refusals, but broad tensor edits appear to impose the largest collateral damage.
• –The strongest value not the uncensoring itself, but the methodology and artifact trail: benchmark cards, KL numbers, tensor overlap, and per-layer breakdowns make the claims checkable.