A user correction should outrank a personalization model

Personalization gets dangerous when the system treats its inference as more authoritative than the person using it.

If a user says “this is too easy,” changes a goal, removes a topic, or asks the system to forget a preference, that correction should take effect immediately. It should not become one more weak signal waiting to be averaged against weeks of behavior.

I think the hierarchy should be explicit:

1. direct user correction

2. current user choice

3. recent observed behavior

4. historical inference

Otherwise “adaptive” quietly turns into an argument with the product. Where have you seen a recommendation system recover well after being corrected?