Scoring integrity lab

# AI quiz partial-credit scoring test

A partial-credit model matrix and exhaustive response-pattern test for detecting guessing rewards, penalty errors, caps, floors, rounding conflicts, and unstable quiz scores.

Published9 October 2026Reading time10 minEvidencePartial-credit model matrix and exhaustive response-pattern testAuthorThe AI Quiz Lab

Test partial credit by defining the scoring rule before authoring the item, enumerating every valid response pattern, and comparing expected points with actual points. Include correct selections, incorrect selections, omissions, caps, floors, rounding, retries, and unanswered states. Publish only when every pattern produces an explainable score and selecting everything cannot earn unintended full credit.

## Key findings, verified 9 October 2026

- Moodle currently supports positive, zero, negative, and partial marks for multiple-choice options.
- McGraw Hill documents three distinct check-all-that-apply modes: as authored, partial credit, and all or nothing.
- NCSBN documents partial-credit scoring for traditional multiple-response items; a separate Select N model awards one point per correct option without an incorrect-option penalty.
- The learner-facing instructions must match the implemented scoring rule.
- Passing a calculation test does not prove fairness, reliability, validity, or suitability for high-stakes use.

## What is partial-credit scoring?

Partial-credit scoring awards some points when a response demonstrates part, but not all, of the intended evidence. It can apply to multi-select, matching, ordering, numeric, short-answer, and multi-part questions. The model is part of the item design: two quizzes can show the same choices yet measure and reward behavior differently.

## Partial-credit model matrix

ModelRuleUseful whenPrimary riskAll or nothingFull points only for the complete correct patternThe response is indivisiblePartial knowledge is invisibleAdditiveAdd points for each correct componentComponents independently show knowledgeSelecting everything may be rewardedAdditive with deductionsAdd correct points and subtract incorrect pointsGuessing needs a costNegative totals or confusing penaltiesProportionalCorrect components divided by required componentsEach component has equal weightWrong selections may be ignoredRubric levelsMap evidence patterns to defined score levelsQuality is holistic or components varyAmbiguous level boundariesCapped partialAdd component points up to a cap unless completePartial knowledge deserves credit but not masteryCap logic may conflict with feedback

## Reproducible scoring test

- State the construct. Explain why partial credit is appropriate.
- Write the rule twice. Use both plain language and formula notation.
- Define boundaries. Record the maximum, minimum, cap, floor, precision, and rounding rule.
- Enumerate responses. For a small item, list every pattern; for a larger item, cover each equivalence class and boundary.
- Include adversarial cases. Test full correct, partial correct, correct plus wrong, all selected, none selected, duplicated, and invalid responses.
- Build an independent oracle. Calculate expected points outside the quiz platform.
- Compare outputs. Submit every case and inspect the score, feedback, and stored response.
- Retest state changes. Repeat after shuffling, retry, reconnect, and export.
- Trace downstream effects. Verify totals, percentages, score bands, branches, and outcomes.
- Require reproduction. Have a second reviewer rebuild the matrix from the published rule.

## Worked example: select two of four

An item has correct options A and C. Each correct choice is worth +1 and each incorrect choice −1, with a floor of zero and a maximum of two. The expected cases include: A+C = 2; A only = 1; A+B = 0 after the deduction; A+B+C = 1; all four = 0; and none = 0. If the implementation assigns two points to all four selections, the negative weights or cap did not execute.

The same item under proportional scoring without deductions could give full credit when every option is selected. That is not a software defect if it matches the authored rule, but it is a design failure when the instructions imply that guessing is penalized.

## Source-bounded QA prompt

Given the item, keyed components, incorrect components, maximum score, minimum score, cap, floor, rounding rule, retry policy, and platform formula below, generate an exhaustive response-pattern matrix. Return expected points, explanation, and boundary class for each pattern. Flag any pattern that rewards an incorrect selection, exceeds the maximum, contradicts the learner instructions, or changes after shuffle, retry, export, or reconnect.

## Publication blockers

- Selecting every option can earn full credit without demonstrating the intended evidence.
- The written rule and implemented formula differ.
- A response can exceed the maximum or fall below the declared minimum.
- Rounding changes a score band or pass/fail result without a stated policy.
- The same evidence earns different points after shuffle, retry, reconnect, or export.
- Feedback says “correct” while the score is partial or negative.
- Learners cannot understand how partial credit is earned.
- The model has not been reviewed for fairness and assessment purpose.

## What this test does not prove

This test verifies implementation against an authored scoring rule. It does not prove that the rule is educationally sound, fair, reliable, accessible, or valid. It does not establish legal or regulatory compliance. High-stakes scoring requires psychometric, subject-matter, accessibility, and governance review.

## Sources, verification date, and corrections

This guide was verified on 9 October 2026 against Moodle's current [multiple-choice scoring documentation](https://docs.moodle.org/en/Multiple_Choice_question_type), McGraw Hill's official [check-all-that-apply scoring modes](https://www.mheducation.com/highered/support/knowledge/connect-check-all-that-apply-question-scoring.html), and NCSBN's current explanations of [partial credit for traditional multiple-response items](https://ncsbn.zendesk.com/hc/en-us/articles/7807555764631-Are-traditional-NCLEX-Multiple-Response-Select-all-that-Apply-SATA-items-scored-with-partial-credits) and [Select N scoring](https://www.ncsbn.org/public-files/NGN_Summer21_Eng_02.pdf). Those sources document platform-specific rules; the model matrix, exhaustive test, worked example, prompt, and blockers are original AI Quiz Lab evidence assets. Report a changed rule or reproducibility problem through the [corrections process](https://bestaiquizgenerator.com/corrections).

## Next checks

Use the [randomization integrity test](https://bestaiquizgenerator.com/blog/ai-quiz-randomization-test) to keep score bindings stable, the [score-band threshold test](https://bestaiquizgenerator.com/blog/ai-quiz-score-band-test) for downstream outcomes, and the [question-quality checklist](https://bestaiquizgenerator.com/blog/how-to-evaluate-ai-generated-quiz-questions) to review each keyed component.

---

Canonical: https://bestaiquizgenerator.com/blog/ai-quiz-partial-credit-test