Uncensoring Methods · March 2026

Ablation vs Heretic vs Obliteratus: one trick, three layers of tooling

Cover image: User provided · pasted-1772480746.png

Ablation, Heretic, and Obliteratus are closely related. The real differences come down to how much tuning, tooling, and workflow each one adds.

Ablation is the core move. Heretic and Obliteratus extend it in different directions.

Refusal behavior lives in identifiable directions inside the model. Edit those directions and behavior moves fast.

From there it turns into tooling. Ablation is the move. Heretic adds search. Obliteratus adds a larger workbench.

The Short Version

• If you want the core idea, start with ablation: identify the refusal direction and suppress it
• If you want that process automated and tuned, you are moving toward Heretic
• If you want a bigger all-in-one platform with benchmarks, UI, steering, and analysis modules, you are moving toward Obliteratus

Where Each Method Acts

Some safety behavior lives in the model and some lives in the serving stack around it.

What Came Before Ablation

People were already changing behavior before ablation. Those older methods were cheaper, easier to reverse, or clearer about the layer they touched.

Ablation stands out because it edits refusal more directly than the methods around it.

Ablation

In the current LLM context, ablation means: find the internal direction most associated with refusal, then suppress or project it out. That is why the method spread so quickly in 2024. It treats refusal as a geometric feature you can edit.

Mechanically, the workflow is simple:

• Cache activations for contrastive prompt sets (harmful vs harmless)
• Compute per-layer mean differences to estimate candidate refusal directions
• Evaluate candidate directions for effect size and collateral damage
• Apply suppression at inference time (hooks) or persist it by orthogonalizing write paths in model weights

That simplicity is the appeal. You do not need a giant retraining run to get visible movement. The tradeoff is drift.

Heretic: ablation with a search loop

Heretic takes that core move and turns it into a search process.

In practice, Heretic:

• derives refusal directions from harmful vs harmless residual comparisons
• applies weighted interventions across layers instead of one global setting
• searches for better compliance-versus-drift tradeoffs
• and keeps evaluating outputs during the loop

The upside is better tuning discipline. The tradeoff is hidden drift.

Obliteratus

Obliteratus turns it into a broader toolset.

In practice, Obliteratus usually:

• collects harmful vs harmless activations and ranks candidate refusal directions
• offers several extraction and projection variants instead of one fixed recipe
• brings metrics like refusal rate, perplexity, coherence, and drift into the main workflow
• and makes iteration easier through UI, CLI, notebooks, and side-by-side comparisons

The upside is lower friction and more visibility. The tradeoff is false confidence.

One Lineage

Bottom Line

• Ablation is the core move: locate refusal-linked structure and suppress it
• Heretic is that move with more automated search, weighting, and tuning
• Obliteratus is that move wrapped in a larger experimentation and benchmarking platform

Same core idea, different workflow, different failure mode.