Why it matters: a specialist that improves the main metric may become less reliable on exceptions, refusals or instructions outside the training distribution.
A strong answer should cover: adaptation capacity, rank choices, data mixture, base-model behaviour and retention tests. Ask the expert to separate demonstrated findings from plausible mechanisms. The learning/forgetting study is a precise starting point, not a universal forecast.
Follow-up: “Which retained capabilities are mandatory for this product, and how would you test them without making the evaluation unmanageably broad?” Require a shortlist of material checks, not an indiscriminate benchmark bundle.
Interview status: Not yet interviewed.