Follow a task through planning, action and verification. Each step makes responsibility and failure visible; the instructor retains the teaching decision.
Prepare a lesson on causal inference
An original, unevaluated design exercise. Use a system to propose alternatives, while keeping the learning goal and source judgment with the instructor. The workflow becomes agentic only where the system chooses its own next actions or tools.
Define the decision students must make
Responsible: Instructor
Choose a narrow goal: explain why a correlation between optional review-session attendance and exam performance does not establish a causal effect. Supply an approved textbook passage and a fictional data table.
Check before continuing: Can students succeed by identifying a plausible alternative explanation, rather than by repeating a slogan?
If the check fails: Rewrite the goal and the task before generating any materials.
Generate competing lesson plans
Responsible: System, within an instructor-defined boundary
Request two lesson sequences, one beginning with a prediction and one with a comparison of rival explanations. Require each activity to name the reasoning students perform. Permit no external search or access to student records.
Check before continuing: Do the plans differ in the intellectual work they ask students to do? Are the invented data clearly marked?
If the check fails: Reject superficial variations and ask for a different sequence, or write it manually.
Test the items against the goal
Responsible: Instructor
Read each proposed question and answer key. Check whether a plausible confound actually precedes or influences both variables. Replace an item that merely asks for the definition of correlation.
Check before continuing: Would an incorrect causal explanation receive credit under the proposed key? Can a student explain a different defensible answer?
If the check fails: Correct the item, key and rationale together. Do not retain an unchecked item because its prose sounds clear.
Try an unfamiliar case
Responsible: Instructor and students
After supported practice, use a fictional campus-gym attendance comparison without the same wording. Ask students to explain their inference independently, under clearly stated conditions.
Check before continuing: Does the new response show application of the reasoning? Interpret it alongside other course evidence.
If the check fails: Revisit the explanation. One unfamiliar task is informative, but it does not isolate the cause of a learning difficulty.
Try the reasoning in a new case
How would you change the sequence if the learning goal were interpreting an interaction, rather than distinguishing correlation from causation?
Draft an assessment without delegating the grade
An original, unevaluated workflow using fictional responses. Item drafting, feedback proposals and decisions about a student’s achievement require different checks.
Specify the inference a score should support
Responsible: Instructor
For an operant-conditioning question, decide whether the aim is defining terms, classifying a case, or explaining the effect of a consequence on future behavior. State permitted help and an accessible alternative.
Check before continuing: Does each scored criterion reflect the intended reasoning, rather than polished wording or unexplained formatting?
If the check fails: Revise the criterion before generating answers or feedback.
Construct contrasting fictional answers
Responsible: System
Ask for three explicitly invented responses: a correct functional explanation, a response confusing negative reinforcement with punishment, and an ambiguous case with too little information about future behavior.
Check before continuing: The instructor checks the classifications and ensures the ambiguous case is recognized as underdetermined.
If the check fails: Repair the conceptual error and annotate why the generated response was misleading.
Compare feedback proposals
Responsible: Instructor, with system suggestions
Apply the proposed rubric to the fictional answers. Invite the system to suggest a revision question for each. Keep all real student work and grade records outside this exercise.
Check before continuing: Does feedback address the student’s stated inference, and can a learner act on it? Is uncertainty admitted where the case lacks evidence?
If the check fails: Replace generic praise or invented diagnoses with a question about the reasoning.
Keep the consequential decision accountable
Responsible: Instructor
Use the exercise to improve the rubric and teaching materials. If considering a real implementation, separately establish institutional permission, data boundaries, moderation and a route to question a decision.
Check before continuing: Can a student understand how their work was judged and obtain a review by a responsible person?
If the check fails: Do not use the workflow to assign grades. A convincing synthetic demonstration is not validation on real students.
Try the reasoning in a new case
Would the same rubric work for an ambiguous social-psychology explanation? Identify the new evidence you would need before adapting it.
Make feedback lead to a revision
An original, unevaluated sequence. The product is a justified revision, rather than an increased volume of comments.
Choose one claim to examine
Responsible: Instructor
Provide an invented paragraph claiming that a large sample removes confounding. Ask the learner to identify the claim and the evidence needed to support it.
Check before continuing: Can the learner distinguish precision from the validity of a causal inference?
If the check fails: Clarify the distinction with a contrasting example before asking for revision.
Ask for a bounded critique
Responsible: System
Give the fictional paragraph and approved methodological explanation to a system. Request two questions that expose the inference, with each question linked to a sentence. No invented citations or rewritten final paragraph.
Check before continuing: Are the questions grounded in the supplied text? Do they leave an intellectual decision for the learner?
If the check fails: Discard unsupported comments and substitute an instructor-written question.
Revise and explain the choice
Responsible: Learner
Have the learner revise the claim and explain which feedback they used, rejected or modified. A defensible rejection of feedback can reveal judgment.
Check before continuing: Does the explanation identify why more participants cannot by themselves rule out a third variable?
If the check fails: Ask for another comparison of competing explanations. Do not equate accepting a comment with understanding it.
Examine later independent reasoning
Responsible: Instructor
Use a different fictional study later, with clear conditions about tool use. Compare the reasoning across tasks without treating prose style or process logs as an integrity score.
Check before continuing: Do several observations support the intended interpretation of learning?
If the check fails: Consider alternative explanations and teaching support. Neither an activity log nor a single answer proves learning or misconduct.
Try the reasoning in a new case
If feedback itself is the object of study, what would you ask students to judge before they revise?