Heretic is an open-source project for modifying and experimenting with language-model behavior through model-level techniques rather than ordinary prompting alone. It is intended for technical users who want to investigate how model responses can be changed by altering or removing learned behavioral directions.

The project provides tooling and workflows for analyzing models and producing modified variants, with an emphasis on reproducible local experimentation. This makes it different from a chat wrapper or prompt library: the work happens closer to model weights and internal representations and therefore requires appropriate machine-learning hardware and technical knowledge.

Heretic is primarily relevant to AI researchers, model hackers, and developers studying alignment, representation engineering, and model behavior. It provides a practical environment for controlled experiments on open-weight models and for understanding how specific behavioral tendencies may be encoded and manipulated inside neural networks.