Knowledge‑editing benchmarks usually certify only local correctness—whether an edited model produces the new fact on near‑edit prompts—without measuring how much of the original fact remains decodable inside the model. We address this by probing residual knowledge directly with a linear trace probe: after editing a fact, we ask whether the original object can still be recovered from the model’s hidden states.\
On GPT‑2‑XL, we applied three mechanistically distinct editors to 50 CounterFact edits: ROME, constrained fine‑tuning, and the memory‑based editor GRACE. All edits achieved 100% generation‑based success, yet the original object remained linearly decodable far above chance (probe accuracy 0.96 for ROME, 0.86 for constrained fine‑tuning, and 0.79 for GRACE, with chance at 0.50). The GRACE result is especially informative: GRACE changes zero base‑model weights, overriding the fact via an external memory, yet the original object is still decodable from the underlying network, indicating that the residual trace cannot be blamed on incomplete weight updates. We interpret this as evidence that editing, even when behaviorally successful, suppresses rather than erases the original association in representational space.\
We also reported a relearning‑savings instrument that behaved unreliably in our setting; we treat this as a negative methodological result rather than supporting evidence. Code and data are released.\
Review