The Tutor Handbook Vol No.0078 | The Support Provenance Check — How a Tutor Interprets Homework When AI, Parents, Peers or Model Answers Helped Produce It Without Treating Supported Work as Independent Evidence

The Tutor Handbook · Volume 0078 · Series ID THB-0078

The Tutor Handbook: Complete series index.

The homework comes back perfect.

The sentences are cleaner than they were in the lesson. The algebra has no missing steps. The Science explanation uses exactly the right causal language. The study plan is beautifully prioritised. The tutor should be pleased.

Then the learner cannot reproduce the same decisions in the next independent attempt.

Nothing dishonest necessarily happened. A parent may have asked useful questions. A classmate may have explained one step. A model answer may have been consulted after the first attempt. A calculator may have checked arithmetic. A spelling tool may have cleaned surface errors. An AI system may have suggested a structure, criticised a draft or generated a worked solution. A teacher may have returned detailed comments. The learner may have used all of these supports appropriately.

The educational problem begins when the returned artefact is treated as though it was produced under conditions that never existed.

The Support Provenance Check asks what assistance contributed to a piece of work, which intellectual decisions still belonged to the learner, and what fresh follow-up is needed before the tutor treats the artefact as evidence of independent capability.

This is not an anti-AI article, an anti-parent article or a demand that students work without legitimate help. Learning often requires support. Feedback, examples, discussion, technology and access tools can make learning possible. The tutor’s responsibility is to preserve the distinction between a better artefact and a stronger learner.

Volume 0073, The Access-Support Boundary, owns the question of which supports are legitimate relative to the skill being tested. This volume owns a different question. The work already exists. Support may have been entirely legitimate. What does the tutor need to know about how the work was produced before using it as evidence?

Quick answer

Ask for provenance at the level that changes interpretation, not for a forensic confession of every small interaction.

For returned work, identify whether the learner received help with understanding the task, choosing the method, generating ideas, organising the response, executing the steps, checking the answer, editing the language or verifying the result. Preserve legitimate access support. Then choose a fresh task that requires the target intellectual operation without the answer-giving assistance whose contribution remains uncertain.

If the learner can reproduce the target decision, the supported artefact and the independent receipt tell a coherent story. If the learner cannot, do not accuse the learner of cheating or dismiss the supported work as worthless. Use the discrepancy diagnostically: the artefact may show what the learner can produce with a certain support and therefore reveal exactly which part has not yet transferred into independent control.

1. Provenance is about the route to the artefact

In ordinary language, provenance means where something came from and how it arrived in its current form. In tutoring, support provenance is the educational route by which a piece of work was produced.

A completed answer can contain contributions from several sources:

• the learner’s prior knowledge;
• school instruction;
• tutor instruction;
• a parent’s prompt;
• a peer’s explanation;
• a worked example;
• a model answer;
• an answer key;
• a calculator or symbolic tool;
• spellcheck, grammar or translation assistance;
• search results;
• AI-generated hints, critiques, outlines or answers;
• teacher comments from a previous attempt;
• legitimate accessibility tools or accommodations.

The tutor does not need to eliminate these influences. The tutor needs to understand which of them changed the intellectual work the learner personally performed.

2. A polished artefact answers a different question from independent performance

A polished essay may demonstrate that a learner can participate in a revision process using feedback and tools. That is a genuine capability. It does not automatically demonstrate that the learner can independently generate the same structure under a new prompt.

A correct Mathematics solution copied after studying a worked example may show that the learner can follow and reproduce a method. It does not automatically demonstrate method selection from an unfamiliar mixed set.

A Science explanation revised after an AI critique may show that the learner can recognise and improve a weak causal link when it is pointed out. It does not automatically demonstrate that the learner would detect the missing link alone.

These are not semantic technicalities. They determine what the tutor should teach next.

3. Support is not one thing

“Did you get help?” is too coarse to be useful. Most learners get help. The important question is what the help did.

A support can change access, orientation, decision, execution, checking or expression.

• Access support helps the learner encounter the task fairly: enlarged text, screen reading, approved notation support or another legitimate accommodation.
• Orientation support clarifies what the task is asking without choosing the answer.
• Decision support suggests a method, evidence source, argument direction or next intellectual move.
• Execution support performs part of the working, calculation, drafting or coding.
• Checking support flags errors, compares with criteria or verifies an answer.
• Expression support improves wording, spelling, formatting or presentation after the underlying idea has been produced.

The same tool can operate in several categories. An AI system can restate a task, generate a method, calculate an answer, critique a paragraph or edit grammar. A parent can ask one clarifying question or dictate an entire response. “Used AI” and “parent helped” therefore tell the tutor very little until the contribution is described more precisely.

4. The target capability determines which provenance matters

Suppose the current target is algebraic representation: turning a word problem into an equation. If a calculator handles arithmetic after the equation is formed, that assistance may not damage the evidence about representation. If an AI assistant proposes the equation, the target operation has been outsourced.

Suppose the target is inference in reading. Keeping the passage visible is normal access to the evidence, not inappropriate support. If a parent says, “Look at the last sentence,” the evidence-selection operation has been narrowed. If a model answer supplies the inference before the learner attempts it, the returned answer cannot be read as an independent inference.

Suppose the target is composition editing. Using spellcheck may be compatible with the target if the tutor is assessing idea organisation. If the target is independent sentence editing, automated correction changes the evidence.

Support provenance is therefore inseparable from a clear learning target.

5. Do not remove legitimate access support in the name of clean evidence

A learner who requires legitimate accessibility support should not have that support withdrawn merely because the tutor wants an “independent” sample. Independence means the learner performs the target intellectual operation without answer-giving assistance; it does not mean removing the conditions that allow fair access to the task.

This distinction is why the Access-Support Boundary must remain active. Provenance records support; it does not automatically classify support as illegitimate.

6. The first attempt is especially valuable

When possible, preserve what the learner produced before substantial feedback or external assistance. The first attempt reveals a starting route. Later versions reveal what changed after support.

This can be as simple as keeping the rough draft, taking a photo of the original working, saving the first paragraph before revision, or asking the learner to mark where they became stuck before consulting help. The purpose is not surveillance. It is to preserve evidence that would otherwise disappear when the final artefact becomes polished.

A first attempt also makes feedback more meaningful. Volume 0072, The Feedback Receipt, asks whether feedback changes later action. Without a visible before-state, the tutor may see a good final answer without knowing which part the learner repaired.

7. Provenance can be recorded lightly

A twelve-year-old should not complete a compliance form after every homework question. The system should be proportional to the educational decision.

Useful lightweight labels include:

• independent first attempt;
• one parent clarification;
• worked example consulted after attempt;
• AI critique used on draft, wording retained only after learner review;
• answer key used for checking;
• peer explanation before reattempt;
• calculator used after method selection;
• teacher comments incorporated;
• access support used throughout.

The label needs only enough detail to change interpretation. If the support does not affect the target capability, recording may be unnecessary. If the learner’s answer depends heavily on outside intellectual work, provenance becomes important.

8. Parent help can be excellent and still change the evidence

Parents often help because they care, because homework is due and because watching a child struggle is difficult. The problem is not parental involvement itself. The problem is losing track of what the child can now do without that involvement.

Compare three forms of help.

A parent asks, “What is the question asking you to find?” This may support orientation while leaving method selection with the learner. A parent says, “Use simultaneous equations.” That changes method-selection evidence. A parent writes the equation and asks the learner to finish the arithmetic. The learner’s final answer may be correct, but most of the representation job was performed externally.

The tutor can value the parent’s support while still arranging a fresh independent check. This prevents the family from being punished for helping and prevents the learner model from being inflated by help that was never meant to prove independence.

9. Peer help has similar layers

Study partners can explain, compare, question and motivate. Peer learning can be useful. Yet a polished group answer does not automatically reveal what each learner personally carried.

If a friend says, “I think the important sentence is paragraph four,” the learner receives an evidence-selection cue. If the friend explains why a particular algebra method fits and the learner then reproduces the route, the work is supported. If the two learners compare independent attempts and each revises their reasoning, the final work contains both personal and collaborative contributions.

The tutor does not need to reconstruct every sentence of conversation. Ask the learner to describe the decisive help: “What could you not do before the discussion that you could do after?” Then create one fresh task that isolates that operation.

10. Model answers create recognition quickly

Model answers are particularly powerful because they can make the destination look obvious after it has been seen.

A learner reads an elegant response and immediately understands why it works. The feeling is real. The knowledge may still be fragile. Recognition of a strong answer and generation of a strong answer are not the same operation.

The existing Sengkang guide How Studying From Model Answers Works owns how to use exemplars as learning resources. The Support Provenance Check asks what the tutor may infer when a later artefact was produced after exposure to the model.

A simple follow-up is to change the prompt while preserving the underlying decision. If the learner can reconstruct the structure under a fresh task, the model has become learning rather than merely a script.

11. Worked examples deserve the same honesty

Worked examples are valuable when learners are acquiring unfamiliar procedures. They reduce unnecessary search and can make hidden reasoning visible. The Sengkang mechanism owner How Worked Examples Work in Learning explains that broader learning job.

After a worked example, however, the tutor should not describe the first near-copy as proof of independent problem solving. It is evidence that the learner can follow or imitate the demonstrated route. The next evidential step is a task that removes some surface similarity or requires the learner to select the method rather than merely execute it.

12. AI makes provenance more important because support can be invisible

Generative AI can provide explanations, examples, outlines, edits, hints, comparisons, questions, translations, summaries and complete answers within seconds. The same screen can therefore act as tutor, editor, calculator, critic and ghostwriter.

The OECD’s Digital Education Outlook 2026, published in January 2026, stresses an important distinction: better task performance with generative AI does not automatically mean learning occurred. The report argues that learning benefits depend on pedagogical purpose and the way tools are used. This is directly relevant to provenance. A high-quality AI-assisted output tells the tutor something about the human-tool system; it does not by itself isolate what the learner can later do when the tool is absent or constrained.

The response should not be blanket prohibition. It should be clarity about the cognitive division of labour.

13. Ask who made the decisive intellectual move

For AI-assisted work, one practical question cuts through much of the confusion:

Who performed the target intellectual operation?

If the learner had to choose the method and AI only checked arithmetic, method-selection evidence may remain strong. If AI chose the method, the learner has not yet demonstrated method selection. If the learner wrote an argument and AI pointed out a missing counterargument, the learner may have demonstrated initial reasoning plus feedback uptake. If AI generated the whole argument and the learner edited wording, the polished essay tells little about independent argument construction.

This is not a moral ranking. Each workflow can be educationally appropriate for a different purpose. The provenance check prevents one purpose from being reported as another.

14. AI literacy includes understanding the boundary of the tool

The OECD and European Commission’s Empowering Learners for the Age of AI, published in June 2026, proposes an AI literacy framework for primary and secondary education that includes engaging with AI critically and responsibly. One practical form of that literacy is being able to say what the tool contributed and what the learner still owns.

A learner who can say, “I generated the first explanation, AI challenged one causal gap, I rewrote it, and I can now explain the mechanism without the tool,” gives the tutor far more useful evidence than either “I used AI” or “I didn’t cheat.”

15. The provenance check should not become a policing ritual

If every returned paragraph triggers an interrogation, learners will learn to hide support rather than use it intelligently.

Build a culture in which support is discussable. The tutor can say:

I am not asking so I can punish you for getting help. I need to know which parts you can already carry and which parts the support was carrying, so I can choose the next useful task.

When the purpose is transparent, provenance becomes metacognition rather than surveillance.

16. Provenance should be proportionate to consequence

A low-stakes practice worksheet may need no formal record. A major claim that the learner is ready to end repair, move to harder work or reduce support deserves clearer provenance.

The greater the consequence of the learner-model update, the more carefully the tutor should understand the conditions that produced the evidence. This principle matches Volume 0076, The Evidence Sample: evidence design should be proportional to the decision, not maximised for its own sake.

17. A discrepancy is a diagnostic gift

Suppose homework is excellent and fresh independent performance is weak. The easy response is accusation: “Someone did this for you.” The better response asks which operation the support made possible.

Perhaps the learner understands once the problem is represented. Perhaps the learner can revise but not generate. Perhaps the learner recognises a strong inference but cannot locate evidence unaided. Perhaps the learner can solve after a method is named but cannot select the method. Perhaps the learner’s ideas are sound and surface language editing creates the polished difference.

The gap between supported and unsupported performance helps locate the next teaching job.

18. Worked case: Alicia and an AI-generated method

This is a fictional teaching case.

Alicia returns a difficult algebra homework set with excellent working. In the next lesson, she hesitates on a simpler problem from the same family.

The tutor asks what support she used. Alicia says she pasted the hard questions into an AI assistant and asked for “a hint.” In practice, the assistant named the method and showed the first equation. Alicia completed the algebra herself.

This is useful provenance. The homework still contains evidence: once the representation and method were supplied, Alicia could execute accurately. It does not support the claim that she can independently represent the situation or select the method.

The tutor does not ban AI. Instead, the next fresh task requires Alicia to write the relationship and choose the method before any tool may be consulted. If she becomes stuck, the agreed AI prompt is changed from “give me a hint” to “ask me one question that helps me inspect my representation without naming the method.” The tutor then checks whether the learner rather than the tool performs the target decision.

19. Worked case: Beatrice and parent-edited English

This is a fictional teaching case.

Beatrice submits a situational-writing draft with unusually concise sentences and precise register. Her parent helped her “tidy the English.”

Rather than treating the draft as unusable, the tutor separates layers. Beatrice chose the content, sequence and audience. Her parent replaced several vague phrases, corrected grammar and removed repetition.

The draft therefore remains evidence about planning and audience decisions, but it is weak evidence about independent sentence control and editing. The tutor asks Beatrice to explain three parent changes and then revise a fresh paragraph without those corrections. If she can apply the same principles, some of the external edit has been converted into learning.

20. Worked case: Ciara and the model answer

This is a fictional teaching case.

Ciara’s Science homework uses textbook-perfect causal phrasing. She says she checked the model answer after writing her own response and then rewrote the whole paragraph to match it more closely.

The tutor asks to see the rough attempt. The original has the correct mechanism but weak precision. That is encouraging. The model answer mainly improved expression. A fresh changed-condition question then tests whether Ciara can reconstruct the same causal relationship in her own words.

If she can, the final homework and fresh response together provide a strong story: the exemplar sharpened expression without replacing the mechanism. If she cannot, the model may have supplied more of the reasoning than the final artefact reveals.

21. Worked case: Emily and the beautiful study plan

This is a fictional teaching case.

Emily arrives with a perfectly balanced revision timetable. Her tutor is impressed until Emily admits that a productivity app generated the schedule after she entered her subjects and deadlines.

The schedule may still be useful. The relevant question is whether Emily can make the planning decisions the current route is meant to transfer to her: identify priorities, protect retrieval, estimate capacity and revise the plan when circumstances change.

The tutor therefore changes the evidence condition. Midweek, one deadline moves and a school task expands. Emily must decide how to update the plan before consulting the app. Her explanation of what she moved, preserved and dropped becomes the independent receipt.

22. Worked case: Faith and peer collaboration

This is a fictional teaching case.

Faith completes a complex problem set with a friend. They discussed each question and compared methods. The final work is accurate.

The tutor does not need to know who said every sentence. Faith identifies two moments where her friend’s explanation changed her route. The tutor then selects one new problem requiring those same decisions. Faith attempts it privately before discussion.

Collaboration is preserved as a learning resource. Independent evidence is preserved as a different condition.

23. Returned work can be evidence of feedback uptake

Supported work is not merely “contaminated evidence.” It can answer useful questions of its own.

A revised draft can show whether the learner understood feedback. A corrected solution can show whether the learner can follow a worked route. An AI-assisted critique can show whether the learner can judge suggestions rather than accept them blindly. A parent-supported plan can show whether the learner can participate in prioritisation with guidance.

The error is not using supported work. The error is asking it to prove unsupported capability.

24. The tutor should create a clean receipt after consequential support

If support performed a target operation, create a later task where the learner performs that operation.

The clean receipt does not need to recreate the original task exactly. It should preserve the intellectual job while changing enough surface detail to prevent copying or recognition from doing the work.

• If AI selected the algebra method, use a fresh mixed problem and ask the learner to select first.
• If a parent reorganised the essay, use a new paragraph and ask the learner to sequence the ideas.
• If a model answer supplied causal language, use a changed Science context and ask for reconstruction.
• If a peer found the evidence, use a new passage and ask for independent evidence selection.
• If an app built the plan, change one real constraint and ask the learner to replan without the app first.

25. The clean receipt should not become punishment

Do not make the fresh task harder merely to “catch” the learner. The aim is to identify what transferred from supported performance into learner control.

A punitive test encourages concealment. A diagnostic check encourages honest support use because learners see that asking for help does not invalidate their work; it simply changes the next evidence needed.

26. School rules still matter

The educational distinction between supported and independent work does not override school rules on collaboration, AI, calculators, translation, plagiarism or assignment assistance. A support can be educationally useful and still be prohibited for a particular assessed task.

Tutors should encourage learners to follow the instructions of the school or assessment authority for the task in question. When current Singapore examination or syllabus rules matter, verify them against current MOE or SEAB sources rather than assuming that ordinary study-tool use predicts what is permitted in formal assessment.

This volume concerns interpretation of learning evidence, not permission to ignore institutional rules.

27. Provenance and privacy

Knowing what support contributed does not require collecting private transcripts, copying personal chat histories or uploading school work into unnecessary systems.

Ask for the educationally relevant description: what kind of help was used and which step changed. If a learner can explain that a tool supplied an outline or a parent corrected grammar, the tutor often has enough information to design the next check.

Provenance should reduce uncertainty without becoming surveillance.

28. Three-student tutorials require provenance before peer comparison

In a small group, one student’s returned homework may be independent, another’s heavily parent-supported and another’s revised with AI. Comparing the artefacts as though they were produced under the same conditions can create false judgments about ability.

Before using homework differences to reclassify the learners, the tutor should know enough about support conditions to interpret them. The next in-class task can then establish a more comparable evidence condition without shaming any learner for how they studied.

This is another reason the group should not become one averaged learner. Shared material can coexist with individual evidence conditions.

29. Parent communication: help is welcome, hidden substitution is the problem

Parents may worry that telling the tutor about help will make the work “not count.” The tutor should explain the distinction clearly.

Please help when help is genuinely needed. I only need to know what kind of help changed the work so I do not mistake supported performance for a skill your child is already carrying alone. I can then test the supported step separately and use your help as part of the learning process rather than pretending it never happened.

This makes families collaborators in evidence quality rather than suspects.

30. Learner communication: provenance is a self-knowledge skill

Ask the learner to answer three questions:

• What did I do before getting help?
• What did the help change?
• What can I now do again without that help?

A learner who can answer these questions is beginning to distinguish exposure, support, revision and ownership. That is useful metacognition regardless of whether the support came from a human or a machine.

31. A Support Provenance card

• Target capability: What intellectual operation are we trying to observe?
• First attempt: What did the learner produce before substantial help?
• Support source: Parent, peer, tutor, model answer, worked example, AI, calculator, editing tool or other?
• Support function: Access, orientation, decision, execution, checking or expression?
• Decisive move: Who performed the target operation?
• Legitimacy: Was the support allowed for this learning or assessment context?
• Artefact claim: What can the supported work honestly show?
• Independent gap: What remains unproven?
• Fresh receipt: What changed-condition task will test the unproven operation?
• Privacy boundary: What information is sufficient without collecting unnecessary personal data?

32. Research foundation: AI-supported output and learning are not identical

The strongest current reason for provenance discipline in AI-supported work is not that AI always harms learning. The evidence and policy landscape is more nuanced. The OECD’s Digital Education Outlook 2026 describes generative AI as capable of supporting learning when used with sound pedagogical intent while warning that successful task completion with AI assistance does not automatically imply underlying learning. The OECD and European Commission’s 2026 AI literacy framework treats critical, responsible engagement with AI as an educational capability in its own right.

Those sources support the distinction between tool-assisted performance and learner capability. They do not establish that every AI-assisted homework task requires the same follow-up, nor do they prove that unaided work is always educationally superior. Provenance is a way to preserve the distinction while allowing the tool to serve legitimate learning purposes.

33. Research foundation: feedback should lead to learner action

The Education Endowment Foundation’s 2026 guidance on feedback and independent learning emphasises specific information linked to goals, reflection and opportunities for learners to act on feedback. That principle matters here because supported revision becomes educationally stronger when the learner can explain and reuse the change rather than merely accept a corrected artefact.

The research supports feedback uptake as part of learning; it does not turn every polished revision into proof of independent production. A later learner-owned attempt remains the cleaner receipt when independence is the claim.

34. Research foundation: tutoring should use data without confusing activity with outcome

The National Student Support Accelerator’s current Tutoring Quality Standards distinguish several quality elements including formative assessment, student progress measures, instructional materials, coaching and programme improvement. Its data-use guidance treats assessment data as information for instructional decisions.

Support provenance is a small-scale interpretation discipline inside that larger idea. Before returned work becomes “data” about what the learner can do, the tutor should know enough about the production conditions to avoid assigning capability to support.

35. Common failure: provenance becomes a morality score

“Independent” can sound virtuous and “supported” can sound inferior. That framing is educationally unhelpful.

New learning is often supported. Expert performance itself can use legitimate tools. The important question is whether the evidence condition matches the claim. A learner can be excellent at collaborating, revising and using tools while still needing to build independent capability in a particular underlying operation.

36. Common failure: support is hidden because adults overreact

If every admission of help leads to punishment, extra worksheets or accusations, learners learn to conceal provenance. The tutor then loses exactly the information needed to teach well.

Reward honest description of support by using it to make the next task more precise.

37. Common failure: all AI use is treated as equivalent

Generating a complete answer, asking for a counterexample, checking a calculation, requesting a quiz, translating an unfamiliar instruction and receiving grammar suggestions are different interactions. They change different learner operations.

Do not make the tool brand the unit of analysis. Make the intellectual contribution the unit of analysis.

38. Common failure: “I understand it now” ends the check

A learner may genuinely understand a model answer, AI explanation or parent’s correction. That matters. Understanding after support is a different state from independent reconstruction later.

Use a fresh receipt when independent capability matters. Do not argue with the learner about whether they “really understood.” Let the next performance answer the question.

39. Common failure: remove every tool to prove independence

Independence should be defined relative to the target and legitimate environment. If the task normally permits a calculator, a dictionary, a screen reader or another access tool, removing it can change the construct being tested.

The goal is not technological purity. It is truthful attribution of the intellectual operation.

40. What this volume does not own

This volume does not define school AI policy, plagiarism rules, examination tool permissions, privacy law or clinical accessibility decisions. It does not claim that AI use causes weaker learning, that parent help is harmful, or that supported work is invalid. It does not replace How AI-Assisted Study Works, How Studying From Model Answers Works, the Access-Support Boundary, or the broader rules of the learner’s school.

It owns one tutor decision: how to interpret a returned artefact when the route that produced it included meaningful help.

Final principle

Do not ask whether the work is “really theirs” as though learning were an ownership dispute.

Ask what the learner did, what the support did, what changed because of the support, and what the learner can now carry into a fresh situation.

The final artefact can be excellent. The support can be legitimate. The learning can be real. The tutor still needs a clean distinction between supported production and learner-owned capability before changing the route.

The Support Provenance Check turns “Who helped?” into a better educational question: “Which part of this work now belongs to the learner strongly enough to travel without the same help?”