Knowledge Assessment System Template

Target a minimum 3:1 ratio (3 bank questions for every 1 on the exam). For high-stakes certification exams, aim for 5:1 or higher. A 50-question certification exam should draw from a bank of 250+ questions. This prevents memorization and ensures form equivalence. For low-stakes practice quizzes, a 2:1 ratio is sufficient since learners retaking the quiz are reinforcing learning, not gaming the system.

Above 0.3 is good. Above 0.5 is excellent. Below 0.2 means the question fails to distinguish between learners who know the material and those who do not. It should be revised or retired. A negative discrimination index (strong learners get it wrong more often than weak learners) means the question is misleading or the keyed answer is wrong. Fix it immediately.

For formative assessments (practice quizzes, module checks): yes, immediately. The feedback is the learning. For summative assessments (graded exams): show after the submission deadline for the cohort. For certification exams: never show correct answers. Showing certification exam answers creates a memorization pipeline that destroys exam validity. Instead, show domain-level performance so learners know where to study.

Layer your defenses. Use in-session writing (browser-locked, timed) rather than take-home essays. Compare writing style against the learner's historical submissions. Ask questions that require personal experience or application to specific scenarios (hard to generate generically). For high-stakes assessments, add an oral follow-up where the learner must explain their written response. Accept that AI is a tool and design assessments that test the ability to evaluate and improve AI output, not just produce text from scratch.

Use adaptive testing when you have a large item bank (500+ calibrated items per domain), need to measure precisely across a wide ability range, and want shorter exams (CAT typically needs 40-60% fewer questions for the same measurement precision). Use fixed-form exams when your item bank is small, you need to compare scores across learners directly, or regulatory requirements demand all learners answer the same questions. Most platforms should start with fixed-form and migrate to CAT after accumulating sufficient item-level performance data.