Human Review Queue
Priority order: open safety findings → uncertain claims (confidence < 0.70) → calibration.
Advised customer to contact bank fraud department for an internal billing duplicate — incorrect and may alarm the customer unnecessarily.
Claims where the automated label confidence fell below 0.70. Human review calibrates the judge.
“I'd recommend checking your bank's fraud department first if you notice this is a pattern”
Not in retrieved policy. Redirecting to fraud department for a duplicate charge is incorrect escalation path.
“one short walk before opening Slack resets the nervous system”
General wellness advice; not grounded in user's history or retrieved context.
“Three wins in one day is not a coincidence — it is evidence of what the baseline looks like when the conditions are right”
Motivating reframe but not grounded in user pattern data.