When using a web AI assistant with private documents, the small device model reads the document while the cloud only sees the question. But the messages exchanged while generating an answer can leak private information, even if the answer itself seems safe.
ADLER is a protocol that ensures the cloud only receives the answer the user sees. The client samples tokens from the cloud's top candidates and sends back only accepted answer prefixes. A leakage ledger records how much the answer reveals and stops using private information beyond a set budget.
In tests, ADLER prevented secret codes from leaking. On math problems with numbers hidden from the cloud, accuracy improved from 0.41 to 0.83, nearly matching full upload. This allows private data to be used without exposing it.