Abliteration
A publicly documented technique that removes refusal behaviour from an aligned model — in minutes, without special access, on consumer hardware. It is the empirical refutation of model-level binding, and the reason L4 cannot be load-bearing.
What It Demonstrates
The technique targets the refusal direction in activation space and ablates it. The capability was never removed; only the expression was suppressed. This is the general form of the failure: behavioural suppression is not capability removal, and the difference is invisible from the outside until someone removes the suppression.
The Rule It Forces
No model-level constraint may be load-bearing for any safety-critical property.
L4 remains useful for precision and defence-in-depth — narrow, high-confidence coverage of named behaviours. It is never the primary binding, because the primary binding must survive an adversary who holds the weights.
Why It Generalises Beyond This Technique
Any constraint implemented in the weights is subject to the same class of attack: fine-tuning, task arithmetic, steering, or a fresh training run. Abliteration is one instance of a general property — weights in the wild is the dominant threat model, not mind-escape.