The scenario: You evaluate your customer support bot in January. Score: 4.2. Ship it! But the world changes — new products launch, user behavior shifts, regulations update — and your system degrades without any code change.
Why it matters: An evaluation is a snapshot, not a guarantee. The number has an expiration date, and you don't know what it is. If you only evaluated once in January, when would you notice the degradation?
The fix: Continuous monitoring, not one-off evaluations. Drift detection alerts. Periodic re-evaluation on fresh production data.
Shock = a sudden event (new product launch, regulation change) that drops scores abruptly. Set month to 0 to disable.
If you had continuous monitoring, you'd catch the drop here.
Simulate 12 Months — Generates a fresh random trajectory for all 4 metrics over 12 months using your slider settings. You'll see stat cards (Jan vs Dec score, total drop, when the alert fires), a line chart of all metrics with your threshold line, a table of real-world drift causes, and key takeaways. Each click re-randomizes — the variation itself is the lesson.
Compare: With vs Without Monitoring — Runs the same simulation, then appends a side-by-side verdict: with monitoring you catch the drop in weeks (when it crosses your threshold); without it, you don't notice until a client complains months later.