Practice feedback and scoring with privacy

Published 2026-10-13 · Version 1.0

Use invented work to distinguish visible evidence from assumptions, inspect provisional criterion annotations, and design a usable revision.

Goal

Rehearse a teacher-controlled chain from a checked task and rubric to feedback, revision, and a fresh check, while keeping real learner information outside the exercise.

Prerequisites

• A teacher-written task and learning goal you can independently check
• Ability to distinguish an observation in a response from an inference about its author
• A fully invented response set unrelated to an actual learner; never a renamed or lightly rewritten real sample
• An institution-approved tool if AI is used; paper-and-colleague practice is equally possible

Proposed time

Proposed planning estimate: 85 minutes across four workflows plus 25 minutes for the practice exercise. Time covers preparation and review, not automated marking throughput.

Ordered route

1. Verify sources and answer keys

Proposed estimate: 20 minutes.

Check the task and expected answer before assessing any invented response. Resolve ambiguities that would make several answers defensible. State what the task cannot reveal.

Checkpoint: A correct, inspectable reference exists; your rubric will not punish an answer for a defect in the question.

2. Calibrate LLM-assisted scoring with synthetic work

Proposed estimate: 30 minutes.

Write criterion-level descriptors and independently annotate three invented anchors. Keep one reference annotation hidden from the model, require exact evidence, and investigate wording or anchor-order sensitivity.

Checkpoint: You can explain disagreements criterion by criterion. Invented evidence, privacy uncertainty, or unexplained score changes trigger a stop, not an averaged score.

3. Write feedback that leads to a revision

Proposed estimate: 20 minutes.

Choose one visible gap in the synthetic response. Draft feedback that asks for a manageable action, reserve time to use it, and write a fresh task with a checked answer.

Checkpoint: The comment does not supply the whole revision or infer effort, ability, emotion, identity, or intent. The new task asks for the same underlying reasoning.

4. Turn formative evidence into a next step

Proposed estimate: 15 minutes.

Use a fictional pattern to compare two short teaching responses. Separate what the evidence shows from two explanations it cannot distinguish. Complete the privacy exercise before considering any separate institutional process.

Checkpoint: An educator selects a provisional teaching response. No model output posts a grade, ranks learners, recommends placement, or becomes a consequential decision.

Finished output

• A checked task, a small criterion rubric, and three wholly invented anchors with teacher reference annotations
• A disagreement record containing actual quoted evidence, any abstention, and the teacher’s reason for keeping or rejecting a suggestion
• One edited feedback comment, a revision opportunity, a checked fresh task, and a privacy stop decision

Self-check

• Are the inputs newly invented rather than transformed real learner records?
• Could a distinctive detail, date, small group, or combination identify someone despite removed names?
• Did you make the reference judgments before seeing the model suggestions?
• Does each proposed criterion annotation rely only on evidence actually present?
• Does a successful synthetic trial remain separate from permission to use actual learner work?
• Are decisions with an authorized educator, with no automatic consequential grading or gradebook posting?

Practice

Open the worked practice exercise. Try it before reading its answer.

Next use

Revise one ambiguous descriptor and test a new invented anchor. Any proposed real-work use requires a separate institution-approved environment and process; this path cannot approve it.

Published October 13, 2026. All time budgets are proposed planning estimates, not measured results. Public sources inform stated principles but do not evaluate or endorse these paths, prompts, worked answers, or exercises. No learning gains, workload savings, or historical classroom outcomes are claimed.