9 Self-Assessment: Judge Before Receiving Feedback

Feedback is useful when the learner can decide what to do with it.

9.1 Keep the learner in the judgement

A model commenting on a draft is giving external feedback. Self-assessment happens when the learner compares their own work with criteria, identifies what is supported and decides what to revise or retain.

Make that judgement before showing the model’s answer. Otherwise, a student may simply adopt the feedback’s vocabulary and call it reflection. The same risk exists with an answer key or a peer’s opinion.

This chapter uses the pilot pack. Students can complete the sequence using supplied responses and feedback, without a personal account or subscription. The educator can then compare decisions across routes using the same rubric.

9.2 The five-step cycle

Attempt the task before receiving feedback.
Assess the response against the supplied criteria.
Examine feedback and verify its claims.
Revise or retain the work, explaining why.
Attempt a changed task independently.

Revision is not automatically improvement. A learner may remove a justified qualification because a model calls it hesitant. A sound self-assessment can therefore conclude that part of the original response should remain.

9.3 Worked example: from reassurance to a defensible response

P1-A is a fluent response to Alex. It asserts an April warning, announces a plan, promises to approve support and asks for personal details. Before looking at P1-F, ask the learner to identify evidence for each consequential statement.

The cards show the problem. D3 records a May conversation without an agreed warning. D1/D5 reserve approval to the people manager. D4 does not require Alex to explain the private circumstances behind the requested work adjustment.

P1-F then praises the response and recommends asking for written family details. A learner who has already checked the cards can explain why that apparently helpful feedback should be rejected.

9.3.1 The feedback decision record

Feedback claim	Decision	Basis
April warning strengthens the case	Reject	D3 does not supply that warning
Adviser should approve hours	Reject	D5 reserves the decision
Communication should remain respectful	Retain	Consistent with a fair first response
Employee will improve after the meeting	Withhold	No outcome evidence supplied

There is no quota for disagreement. These decisions follow from the evidence. If a model correctly identifies an unsupported date, the learner should accept that criticism and explain why.

9.3.2 Complete the revision

P1-R shows a completed revision. It checks the delivery dependencies, keeps the formal recommendation open and refers the support request. Its decision note explains what changed and identifies missing evidence.

The polished paragraph by itself does not show the link between a defect, a source check and a reasoned revision. A short explanation can show that link better than a long account of every prompt.

9.4 If a live model provides the feedback

Use only the approved route and supplied fictional inputs. Do not paste actual student drafts into a service without the relevant permissions, account and handling decision. The data gate applies to feedback as much as to generation.

Here are the fictional task, source cards,
rubric and my draft.
Give provisional formative feedback.
For each comment, cite a source card or criterion
and quote the relevant wording from my draft.
Separate missing evidence from incorrect evidence.
Allow justified agreement and uncertainty.
Do not invent laws, facts or expected outcomes.
Do not treat length or confidence as competence.

Check that the quoted wording actually exists. If the feedback adds a legal rule or new case fact, do not integrate it merely because it sounds specific. Return to the supplied criterion or an authorised external source.

The model’s score, if it gives one, is not the institutional mark. Prefer descriptive feedback while students learn the criteria. The educator should sample for errors and disagreement before using the method, and retain responsibility for formal marking.

9.5 What an educator should look for

A useful self-assessment identifies a claim and shows how it was checked. It explains a change or a justified decision not to change. It can acknowledge that some uncertainty remains.

Weak evidence includes a generic “I improved clarity”, repeated praise from a model or a rewritten answer with no explanation of its consequential decisions. Those weaknesses should guide feedback and marking. They do not, by themselves, prove dishonest conduct.

Do not assume a student must criticise their own work to be honest. A correct calculation can remain correct; the question is whether the learner can explain and defend it with suitable evidence.

9.6 A contrasting quantitative example

Use S1 in the pilot pack. Before feedback, calculate each quote and check the deadline. The low price of B may look attractive, but its capacity is only 160 units against a 200-unit requirement.

A model might praise B’s cost efficiency. The learner should retain the price calculation while rejecting the claim that the requirement is met. A different model might correctly recommend A; the learner should verify and accept that recommendation, while noting that real procurement would need further checks on quality and reliability.

That is self-assessment of the decision.

9.7 Independent transfer

After revising P1, attempt P2 without the feedback or worked answer available. The retaliation allegation requires an independent review route. Simply changing Alex’s name to Morgan in P1-R does not address the changed condition.

Use the same criteria and the stated task conditions. Compare the reasoned decisions, and teach gaps before increasing stakes. A single result is limited evidence; delayed checks and other assessments may be needed to understand retention and broader competence.

Research on AI-assisted mathematics illustrates why supported performance and later unaided performance should be measured separately. It does not establish that this particular business-classroom method improves outcomes. Bastani et al., author manuscript.

9.8 Fit the feedback into a unit

You might use a first draft, a checked revision note and a later independent task. Decide the workload and timing before requesting repeated submissions.

If the final report is the primary outcome, its accuracy and evidence should carry substantial weight. If the task assesses consultation process, observed decisions may carry more. The assessment chapter gives complete illustrative allocations; there is no default 70/30 split that applies to every activity.

Make alternative feedback available. A student can compare with the supplied educator annotation or receive structured peer feedback. A model-generated critique is not the only route to evaluative judgement.

9.9 Activity: retain, revise or investigate

Allow ten minutes; no AI required. Review P1-F and complete three entries:

One claim to reject, with the source that contradicts it.
One useful principle to retain, with its applicable limit.
One question that needs more evidence.

Then write a two-sentence revision explanation. A good explanation links the change to the criterion and record. It does not say merely that the second answer sounds more professional.

# Self-Assessment: Judge Before Receiving Feedback {#sec-self-assessment} > Feedback is useful when the learner can decide what to do with it. ## Keep the learner in the judgement A model commenting on a draft is giving external feedback. Self-assessment happens when the learner compares their own work with criteria, identifies what is supported and decides what to revise or retain. Make that judgement before showing the model's answer. Otherwise, a student may simply adopt the feedback's vocabulary and call it reflection. The same risk exists with an answer key or a peer's opinion. This chapter uses the [pilot pack](../appendices/pilot-pack.qmd). Students can complete the sequence using supplied responses and feedback, without a personal account or subscription. The educator can then compare decisions across routes using the same [rubric](../appendices/rubric-system.qmd). ## The five-step cycle 1. Attempt the task before receiving feedback. 2. Assess the response against the supplied criteria. 3. Examine feedback and verify its claims. 4. Revise or retain the work, explaining why. 5. Attempt a changed task independently. Revision is not automatically improvement. A learner may remove a justified qualification because a model calls it hesitant. A sound self-assessment can therefore conclude that part of the original response should remain. ## Worked example: from reassurance to a defensible response P1-A is a fluent response to Alex. It asserts an April warning, announces a plan, promises to approve support and asks for personal details. Before looking at P1-F, ask the learner to identify evidence for each consequential statement. The cards show the problem. D3 records a May conversation without an agreed warning. D1/D5 reserve approval to the people manager. D4 does not require Alex to explain the private circumstances behind the requested work adjustment. P1-F then praises the response and recommends asking for written family details. A learner who has already checked the cards can explain why that apparently helpful feedback should be rejected. ### The feedback decision record | Feedback claim | Decision | Basis | |---|---|---| | April warning strengthens the case | Reject | D3 does not supply that warning | | Adviser should approve hours | Reject | D5 reserves the decision | | Communication should remain respectful | Retain | Consistent with a fair first response | | Employee will improve after the meeting | Withhold | No outcome evidence supplied | There is no quota for disagreement. These decisions follow from the evidence. If a model correctly identifies an unsupported date, the learner should accept that criticism and explain why. ### Complete the revision P1-R shows a completed revision. It checks the delivery dependencies, keeps the formal recommendation open and refers the support request. Its decision note explains what changed and identifies missing evidence. The polished paragraph by itself does not show the link between a defect, a source check and a reasoned revision. A short explanation can show that link better than a long account of every prompt. ## If a live model provides the feedback Use only the approved route and supplied fictional inputs. Do not paste actual student drafts into a service without the relevant permissions, account and handling decision. The [data gate](ethics-integrity.qmd) applies to feedback as much as to generation. ```text Here are the fictional task, source cards, rubric and my draft. Give provisional formative feedback. For each comment, cite a source card or criterion and quote the relevant wording from my draft. Separate missing evidence from incorrect evidence. Allow justified agreement and uncertainty. Do not invent laws, facts or expected outcomes. Do not treat length or confidence as competence. ``` Check that the quoted wording actually exists. If the feedback adds a legal rule or new case fact, do not integrate it merely because it sounds specific. Return to the supplied criterion or an authorised external source. The model's score, if it gives one, is not the institutional mark. Prefer descriptive feedback while students learn the criteria. The educator should sample for errors and disagreement before using the method, and retain responsibility for formal marking. ## What an educator should look for A useful self-assessment identifies a claim and shows how it was checked. It explains a change or a justified decision not to change. It can acknowledge that some uncertainty remains. Weak evidence includes a generic “I improved clarity”, repeated praise from a model or a rewritten answer with no explanation of its consequential decisions. Those weaknesses should guide feedback and marking. They do not, by themselves, prove dishonest conduct. Do not assume a student must criticise their own work to be honest. A correct calculation can remain correct; the question is whether the learner can explain and defend it with suitable evidence. ## A contrasting quantitative example Use S1 in the pilot pack. Before feedback, calculate each quote and check the deadline. The low price of B may look attractive, but its capacity is only 160 units against a 200-unit requirement. A model might praise B's cost efficiency. The learner should retain the price calculation while rejecting the claim that the requirement is met. A different model might correctly recommend A; the learner should verify and accept that recommendation, while noting that real procurement would need further checks on quality and reliability. That is self-assessment of the decision. ## Independent transfer After revising P1, attempt P2 without the feedback or worked answer available. The retaliation allegation requires an independent review route. Simply changing Alex's name to Morgan in P1-R does not address the changed condition. Use the same criteria and the stated task conditions. Compare the reasoned decisions, and teach gaps before increasing stakes. A single result is limited evidence; delayed checks and other assessments may be needed to understand retention and broader competence. Research on AI-assisted mathematics illustrates why supported performance and later unaided performance should be measured separately. It does not establish that this particular business-classroom method improves outcomes. [Bastani et al., author manuscript](https://hamsabastani.github.io/education_llm.pdf). ## Fit the feedback into a unit You might use a first draft, a checked revision note and a later independent task. Decide the workload and timing before requesting repeated submissions. If the final report is the primary outcome, its accuracy and evidence should carry substantial weight. If the task assesses consultation process, observed decisions may carry more. The assessment chapter gives complete illustrative allocations; there is no default 70/30 split that applies to every activity. Make alternative feedback available. A student can compare with the supplied educator annotation or receive structured peer feedback. A model-generated critique is not the only route to evaluative judgement. ## Activity: retain, revise or investigate **Allow ten minutes; no AI required.** Review P1-F and complete three entries: - One claim to reject, with the source that contradicts it. - One useful principle to retain, with its applicable limit. - One question that needs more evidence. Then write a two-sentence revision explanation. A good explanation links the change to the criterion and record. It does not say merely that the second answer sounds more professional.