Criteria-Aligned Exemplar Development

Workbook-grounded reconstruction. Criteria-aligned exemplars can make assessment expectations visible, support calibration and help students connect criteria to evidence and judgement. Their form, source, audience and release timing matter because one polished answer may narrow legitimate approaches or undermine a live assessment.

Proposed. Developing a defensible exemplar requires alignment with task, criteria, level, conditions and evidence without implying one correct response. Generated answers may invent sources, exceed the expected standard, hide assumptions, encode irrelevant stylistic norms or undermine integrity when access and release are unclear.

Proposed. Develop a criteria-aligned exemplar or small exemplar set for a defined teaching, feedback or moderation purpose. Select the least complete form that serves the purpose; draft against the actual task, criteria, level, conditions and authorised sources; annotate the relationship between evidence and quality; and preserve legitimate alternatives. Verify all content, disclose relevant GenAI involvement and test level, accessibility, security and unintended signalling. Require educator approval of access, release, reuse and retirement, and never use resemblance to the exemplar as an automatic grading rule.

• A clearly purposed exemplar or exemplar set aligned with task and criteria.
• Annotations showing how evidence and decisions relate to quality judgements.
• Verified sources, disclosed generation and explicit limitations.
• An approved plan for access, timing, reuse and evaluation.
• Visible legitimate variation and an explicit prohibition on automatic grading by exemplar resemblance.

Educators define purpose, quality, level and legitimate variation. Disciplinary reviewers verify content; assessment leaders decide access, timing and reuse; accessibility reviewers and students may test interpretation. GenAI structures, drafts and annotates provisionally. People decide whether the exemplar is accurate, fair, secure and educationally useful, and assessors never surrender marking judgement to similarity with it.

Reconstructed risks

• The model contains fabricated evidence, citations or disciplinary errors.
• Students treat one response as the only acceptable form of success.
• The exemplar is set at an unrealistic level or reveals secure assessment content.
• Generated wording is reused without understanding, weakening learning evidence.
• The exemplar encodes cultural or linguistic norms unrelated to the criteria.
• Student work is reused without appropriate permission, context or privacy controls.
• Assessors reward similarity to the exemplar instead of applying the criteria to evidence.

• Minimum viable implementation: select one criterion, generate an annotated synthetic excerpt under the actual task conditions, verify it manually and test whether students or assessors understand both the quality and its limitations.
• Robust implementation: add a structured purpose and release brief, multiple exemplar forms, source records, level checks, GenAI disclosure, accessibility review, integrity stress testing, versioning and retirement rules.
• Low-code implementation: connect an approved assessment store, GenAI drafting workspace and annotation template while keeping access status and educator approval visible.
• Local or privacy-preserving implementation: keep pre-release exemplars and any authorised student work in controlled storage, minimise identifiers and prefer synthetic examples when reuse permission is absent.
• Speculative implementation: generate contrasting exemplars for several valid approaches or performance levels, while requiring source verification, moderation and explicit protection against automated grading by resemblance.