Subliminal Unlearning
Controlled transfer and mechanism audits — does an unlearning method change the resource surface for recovering held-out facts from behaviorally floor-equivalent models?
Subliminal Unlearning is the rigorously audited continuation of Ghosts in the Model, asking whether unlearning methods change the resource surface for recovering held-out facts from behaviorally floor-equivalent models.
- Designed a checksum-sealed 504-fact experiment with exposure-balanced pilots and held-out gates.
- A forensic audit caught a target-exposure confound before it contaminated claims — the kind of methodological discipline that separates reliable research from wishful thinking.
- A rank-1 mechanism explained 86.9% of paired activation energy yet failed output reconstruction by roughly 350×, decisively rejecting the hypothesis.
- Every claim is code-replayed on CPU and checksum-sealed; runs are marked terminal and “do-not-rerun” to prevent cherry-picking.
- Qualification V3 teacher passed at update 504 — 43,416 raw records independently replayed on CPU.