Skip to content
Validity typeInternal
Pass conditionThe mechanism disappears when the capability is removed
Evidence familyTraining (interventional)
Minimum reportingMethod of capability removal, mechanism measurement before and after, behavioral confirmation
Common failure modeNever testing what happens to the mechanism when the capability is trained away

Offset coupling is the converse of onset coupling (I11). It asks: if the capability is removed — by fine-tuning, by ablation, by catastrophic forgetting — does the mechanism go away too? A mechanism that persists after the capability is gone is not tightly coupled to it.

Satisfied when:

  1. The capability is removed. Fine-tuning on a different task, training past the phase transition, or targeted unlearning eliminates the behavior.
  2. The mechanism is measured after removal. The structural or activation signature of the mechanism is checked in the modified model.
  3. The mechanism is absent or degraded. If the mechanism persists unchanged after the capability is gone, offset coupling fails.

The induction heads (token copying) case study is capped partly by I12. Induction heads have been shown to appear when in-context learning appears (onset coupling), but the converse — fine-tuning away the copying behavior and checking whether induction heads disappear — has not been tested. Without this, the mechanism could be an architectural feature that persists regardless of the capability.

Part of the developmental coupling block (I11–I12). Required for Validated tier. Offset coupling is the top-down leg of Craver’s (2007) mutual manipulability criterion: if the capacity goes, its mechanism should go with it.