Skip to content

interlens.interp.patching

interlens.interp.patching

Patch dataclass 

Patch(
    activations: Tensor,
    layer: int,
    positions: tuple[int, ...],
)

Activation patching: overwrite a decoder layer's residual at specific token positions with saved activations (captured from another run/branch).

This is the cross-branch causal-tracing primitive: capture activations at turn N in one branch (via ActivationCache), then inject them at the aligned positions of another branch's forward. The harness owns this because only it knows the turn/position correspondence between branches.

P2 applies the patch on the (single) prompt forward — positions index into the prompt sequence. Aligning positions across branches is the caller's responsibility; Patch just performs the overwrite.