In two-hop reasoning, a model finds an intermediate entity and then uses it. If you rewrite that intermediate representation with a J-lens or R-lens edit, does the rest of the computation actually use what you wrote?
Swapping the coordinates layer after layer turned out to cancel itself. Clamping them to the values from a clean run instead raised the target-answer flip rate on Qwen3.5-9B from 10% to 41%. From there the work separates three things: directly pushing the answer, rewriting the content, and the conditions under which downstream layers read it.
The clamp became a feature in TransformerLens, so other researchers can use and check it.