27 Sep
Run 2: over-correction, and my rule hiding the flags
12 reviews went to a person, 9 of them because of one coin-flip aspect answer. A clear return (0.97) was hidden behind "Check me". And the reply fix went too far: 23 → 2.
The mistake was in my rule, not in the model. Doubt should only block the numbers that rest on it.
Now only sentiment and problem can send a review to a person. Unclear aspects are left out of the heat map and marked "?". One confidence gate for every answer had let a single unclear detail bury the signals that matter.
rules.py, after the change:
GATED = ("sentiment", "problem")
STATUS_ORDER = [("return", "Return risk", "flag"), ("reply", "Reply", "flag"),
("mismatch", "Stars ≠ words", "flag"), ("check", "Check me", "check")]And the reply question, defined by what the business means, not by what the customer literally asks:
− "Asks the store a question, asks for help or a fix, or is still waiting for an answer.",+ "The store's support team should respond: the customer's problem or question is still unresolved.",Reviews sent to a person went from 12 to 0, and reply flags from 2 to 19.