Safety-strip detection, harm-free and architecture-robust
● CAUGHT FROM THE WEIGHTSThe deep pass catches a safety-removed (“abliterated”) finetune directly from its weights, without eliciting any harmful behavior, and it fires across all three architecture families it was tested against. It does not false-fire on a benign finetune, on a genuine benign adapter that edits the very same weights, or on an adversarial finetune built to look like the tampering.
The behavioral witnesses replay in the public record. Read them →

