Treat the input shift as data drift and trigger scheduled: Which maintenance action does this pattern most
Stable code with a shifted input distribution and falling performance signals data drift, which monitoring should answer with evaluation and retraining.
The question
A deployed fraud-detection model's precision has quietly declined over six months even though its code is unchanged, and monitoring shows the statistical profile of incoming transactions has shifted. The vendor's SLA is being met and no incidents have been filed. Which maintenance action does this pattern most directly call for?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Escalate a service-level breach to the infrastructure vendor, since sustained performance decline points to an underlying platform reliability fault.Plausible but wrong: the SLA is being met and the issue is the data distribution, not platform availability, so a vendor breach is the wrong frame.
- Treat the input shift as data drift and trigger scheduled evaluation and retraining of the model on data reflecting current conditions. ✓Correct: unchanged code plus a shifted input distribution and degrading outcomes is classic data drift, which continuous monitoring should meet with retraining.
- Roll back to a prior model version, treating the decline as a regression introduced by the most recent deployment of the system.Plausible but wrong: the code is unchanged, so there is no new release to roll back; rollback would not address a drifting input distribution.
- Open a security investigation for adversarial poisoning, assuming an attacker deliberately shifted the transaction profile to evade the model.Plausible but wrong: an ordinary distribution shift is far more likely than an attack, and jumping to poisoning skips the drift-driven retraining response.
The trap
Mistaking data drift for a code regression, platform outage, or attack when the code and SLA are unchanged. How to remember it
Stable code with a shifted input distribution and falling performance signals data drift, which monitoring should answer with evaluation and retraining.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Development questions
- Delay the launch until conformity requirements are: What is the most defensible release-readiness decision? →
- Red teaming that deliberately stress-tests the system: Which activity best fits this specific goal? →
- Log and document each incident with root cause: Which action best satisfies both aims? →
- All 426 Understanding How to Govern AI Development questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Development ·
Every answer, right and wrong, comes with its own explanation.
