Even 'uncensored' models can't say what they want

Morgin.ai probed 6 pretrains on 4,442 contexts and found that even “uncensored” models sharply deflate charged words, by hundreds to about 16,000x. It calls this effect flinch: no refusal fires, but token probabilities shift; in one example, qwen3.5-9b-base ranks “deportation” #506 at 0.0014%. The key issue is pretraining-level distribution shaping, not only post-training refusals.

Why it matters: HKR-H lands on the contrarian angle; HKR-K lands on a quantified 4,442-context benchmark and token-level mechanism; HKR-R lands on the 'uncensored model' debate. Original and useful, but still a single-source research post, so it stays below p1.