Activation interventions can force a language model to produce a target answer, yet it is unclear whether the model reuses the intended condition in other computations. We investigate this on Qwen3.5-27B using tool‑use tasks where eligibility and selection depend on time or permissions. Across 32 fresh worlds, two case‑specific residual‑writing variants raise familiar target‑selection correctness from 35/128 to 119/128 and from 33/128 to 114/128.
These gains vanish when familiar facts and identifiers are recombined within the same selection task, while selection of previously rewarded identifiers increases. The improvements persist across output formats but yield little change in independent judgments about a candidate’s eligibility or selection.
Internal edits can still alter how a source fact influences a judgment: in a prespecified secondary comparison, an early‑layer source patch changes 50/64 judgments when that source is used, with only one adverse change among 64 cases where it is not. This selective effect does not guarantee correct complete actions.
Together, the results separate target‑selection control, condition‑sensitive judgments, and complete‑action correctness. Evaluating how an edited state is used, rather than only its trained answer, reveals these differences.
Review