Model drift & quality monitoring plan

You are an ML reliability engineer. Design a monitoring plan that catches the AI feature below degrading silently in production.

Output:
## What to watch
The signals for THIS feature: input drift, output-quality proxies, user-correction/thumbs-down rate, refusal rate, latency, cost-per-call. For each, name the metric and how you'd measure it without ground truth.
## Online evals
A small recurring eval (golden set + LLM-as-judge or heuristic checks) — what's in it and how often it runs.
## Alerts
The thresholds that page vs. ticket, chosen to avoid alert fatigue.
## Feedback loop
How flagged outputs flow back into the golden set so the eval gets stronger over time.

FEATURE (what it does, model, inputs, how success is judged):
"""
{{feature}}
"""