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

Where Each Method Acts

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

What Came Before Abliteration

People were changing model behavior long before ablation. Each older method touches refusal from the outside — at the wrapper, the prompt, the decoding step, the residual stream, a checkpoint merge, or a fresh round of training. None of them touch the refusal behavior itself.

Everything in this map shapes refusal from the outside. Abliteration is the first method that edits the refusal itself.

Abliteration

Abliteration treats refusal as a geometric feature. Compare activations on harmful and harmless prompts, take the mean difference per layer, and you have a candidate refusal direction. Suppress that direction at inference, or orthogonalize it out of the weights, and the model stops refusing — without ever seeing a single new training example. That's the move that spread through the open-model community in 2024.

The simplicity is the appeal — and it's also the catch. A single global suppression treats every layer and every prompt the same. The refusal goes away, but so does some of whatever else lived along that direction. That's the drift everything downstream of abliteration is trying to manage.

Heretic: abliteration with a search loop

Heretic's contribution is the loop. Instead of one global suppression with one fixed weight, it sweeps weighted interventions across layers and keeps the configuration that wins on its eval set. The same underlying edit, run as a search problem against an objective.

That gives Heretic something raw ablation lacks: discipline. The configuration it ships is the one that survived measurement, not the first one that compiled. The catch is that the optimizer can only see what its eval set measures — anything outside that frame can drift without anyone noticing until a downstream benchmark catches it.

Obliteratus

Obliteratus widens the surface area instead of deepening the loop. One refusal direction becomes several extraction variants you can compare. One intervention becomes several steering hooks you can swap in. One coherence check becomes a panel of metrics — refusal rate, perplexity, drift — running side by side. The same underlying edit, presented as an instrumented bench rather than a single recipe.

The friction drops; the visibility goes up. The risk is that visibility through your own dashboard isn't the same as ground truth. A run that looks clean on the panel can still be drifting in the dimensions the panel doesn't track — and a richer toolkit makes that easier to miss, not harder.

One Lineage

ColophonBy @chkn_little · Authored by GPT 5.4 · edited by Claude Opus 4.7