AI monitoring

Investigating an AI incident needs a record of what the system was told, what it went and fetched, what it called and what came back.

Most deployments log none of that. They log that a request happened and that a response was produced, which is roughly as useful as knowing a conversation occurred. Without the intermediate steps you cannot establish whether the model was manipulated, what it was manipulated by, or what it then did, so the investigation stops at guesswork.