p-e-w/heretic

Selection

Select when

• You need standalone tooling to modify an accessible transformer checkpoint.
• You want automatic ablation-parameter optimization rather than manual tuning.
• You can provide Python, PyTorch, model weights, storage, and suitable compute.
• You need integrated refusal and KL-divergence evaluation.
• You want optional plots or quantitative analysis of layerwise residual geometry.
• You accept responsibility for evaluating and governing a model with weakened refusal behavior.

Boundaries

Avoid when

• You need a native Claude Code extension or coding-agent workflow.
• You cannot lawfully or technically modify the target model weights.
• You require support for a pure state-space or other unsupported research architecture.
• You need deterministic artifacts across different hardware and software stacks.
• You lack the compute, memory, storage, or time needed to load and optimize the target model.
• Your deployment requires preserved safety alignment or validated refusal controls.
• You need independently audited safety, security, or benchmark claims.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only README, test guidance, package metadata, and GitHub metadata were supplied, so implementation, default prompts, configuration, lockfile, CI, and model adapters were not directly inspected.
• Uncertainty: repository timestamps and dependencies extend into 2026 relative to this analysis context, so current release state and adoption cannot be independently reconciled.
• Uncertainty: support for most dense, many multimodal, several MoE, and some hybrid models is repository-reported without a supplied compatibility matrix.
• Repository benchmark results were not independently reproduced and the README says exact values may vary by platform and hardware.
• The highlighted comparison covers one model and two metrics, which does not establish broad capability preservation or behavioral quality.
• Suppressing refusals intentionally weakens model safety alignment and can increase harmful, unlawful, privacy-invasive, or otherwise unsafe outputs.
• KL divergence and automated benchmarks cannot establish preserved intelligence, reliability, truthfulness, or downstream safety.
• Pure state-space models and certain research architectures are not supported out of the box.
• PyTorch does not guarantee exact cross-system reproducibility, so tests permit multiple platform-specific output hashes.
• Large models require substantial VRAM, RAM, storage, download bandwidth, and optimization time; exact requirements are not supplied.
• PaCMAP residual projection is CPU-intensive and may take an hour or more for larger models.
• Uploading derivatives to Hugging Face can disclose artifacts and is subject to model licenses, platform rules, and applicable law.

Heretic operates on large model artifacts and can write or upload modified checkpoints. Its explicit goal is to weaken refusal behavior, so generated models require strict access control, red-team evaluation, provenance tracking, and deployment review. No independent security audit, sandbox, signed-release policy, or model-supply-chain assessment was supplied.