This observation covers the approximately 24-hour period from 7:48 am on 1 November to 7:48 am on 2 November 2026 in Australia/Sydney. One formally released and verifiable technical change is included.
Heretic 2.0 turns a refusal-removal tool into a pluggable model-modification framework
The open-source Heretic project released version 2.0.0 on 1 November, and PyPI made the heretic-llm 2.0.0 source distribution and wheel available the same day. Earlier releases centred on automatically searching for directional-ablation parameters that reduce refusals on selected prompts while using KL divergence to constrain changes to the model. Version 2.0 separates scoring and model-modification methods into plugin interfaces: scorer plugins define objectives to minimise or maximise, while modifier plugins apply changes to the model. The bundled scorers now include a benchmark scorer and a BERT-classifier scorer. Arbitrary-Rank Ablation, or ARA, is the default modifier, while classic directional ablation remains selectable.
The release notes also confirm headless operation and end-to-end tests, support for Hyper-Connections and flattened residual streams, improved reasoning-prefix detection, and fixes affecting ARA and retained LoRA state. Headless operation allows batch scripts to modify and test models without interactive confirmation, changing the automation conditions from the 1.x series. The project requires Python 3.10 or later and is distributed under AGPL-3.0-or-later. PyPI lists two distribution files for this version but no build-provenance metadata.
The compatibility boundary has also changed: configuration files made for 1.x do not work directly with 2.0 and must be migrated. The previous research features were removed, with the release notes naming Lophius as the intended replacement. The repository and PyPI confirm the version, interface changes and distribution status. The developer's claims about ARA performance and efficiency have not received a public independent reproduction, and the release materials do not provide a complete 2.0-versus-1.x comparative evaluation. Third-party reporting confirms the release and configuration break but does not include an independent model-modification experiment.
Sources
GitHub, Release v2.0.0 · p-e-w/heretic, 1 November 2026, https://github.com/p-e-w/heretic/releases/tag/v2.0.0
Python Package Index, heretic-llm 2.0.0, 1 November 2026, https://pypi.org/project/heretic-llm/2.0.0/
Startup Fortune, Heretic 2.0 makes stripping AI safety guardrails even easier to automate, 1 November 2026, https://startupfortune.com/heretic-20-makes-stripping-ai-safety-guardrails-even-easier-to-automate/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.