Heretic automates abliteration and reports lower capability loss than hand-tuned versions

Sources: primary, project, discussion
Summary: Heretic is not a new release. The project dates to 2025 and reached the Hacker News front page on 2026-09-21 at 196 points. Its README reports 3 of 100 refusals on gemma-3-12b-it at a KL divergence of 0.16 from the original weights, against 0.45 and 1.04 for two hand-made abliterations measured at the same refusal count, and ships an evaluate command that reproduces those numbers, which are author-reported.
Why it matters: Removing refusal behaviour from an open-weight model is now one command with a measured quality cost, which bounds what a safeguard evaluation of published weights can claim once the weights are out.