Case context: You are building an automatic moderation model for a community app that must block prohibited uploads. The current score gives the same importance to an ordinary labeling mistake, such as calling a mountain photo a beach photo, and a serious mistake, such as allowing a clearly forbidden file to pass.

Question: Using the idea of redesigning the evaluation score to reflect costly failures, what should the team change?

Sample answer: The team should revise the evaluation score so that passing through a prohibited upload is treated as far more costly than an ordinary mistake. Rather than counting every error the same way, the metric should assign a large penalty to cases where disallowed content is marked as acceptable. That change will make such failures dominate the score and keep models with those mistakes from being chosen.

Key points:

• Recognize that the current score does not distinguish between mild and serious mistakes.
• Add a much larger penalty for allowing prohibited uploads to pass.
• Explain that the revised score will steer selection away from models that miss these critical errors.

Rubric: The answer must state that the evaluation score needs to be changed, specify that a strong penalty should be applied to the serious error of allowing prohibited uploads through, and explain that this penalty affects model choice by rejecting models that make that mistake.