A quality gate for AI-generated content: 50,000 items checked, about 1 in 10 caught

Context

The network teaches core academics through personalized AI instruction. Its curriculum generator produces math questions, worked examples and figures across about 400 Common Core math standards, and items reach students without a person checking each one.

The problem

Generated items fail in ways a text check can't see: labels too small to read, figure elements overlapping, broken images, raw LaTeX in the rendered page, a worked example that doesn't match its question. Reviewing every item by hand doesn't scale.

What was built

• A library of about 340 parameterized functions that draw the math figures students see, kept legible and correct across every parameter combination the generator produces
• A deterministic gate that renders each item in a headless browser and checks for unreadable text, overlapping elements, broken images, raw LaTeX and worked examples inconsistent with their question
• Automatic deactivation or regeneration of failing items
• Both components scored continuously inside the platform's closed-loop optimization system

The client is not named because the work was done under a confidentiality agreement.