How to Store Original Student Evidence Beside an AI Score

A score records a judgement; it does not preserve the work being judged. When introducing AI-assisted assessment, treat the original submission, the AI output and the teacher’s decision as separate records linked by one submission ID.

This guide explains how to store original student evidence beside an AI score using an assessment-evidence register, versioned records and a retrieval drill. The workflow is an editorial recommendation—not an Ofqual-mandated schema, a compliance guarantee or a verified SubSchool evidence-storage feature.

Why original evidence must remain retrievable

On 14 October 2026, Ofqual announced a £150,000 fine against ABRSM over assessment-evidence retention failures. The announcement describes missing recordings and marking records that affected reviews of marking and a malpractice investigation. In one incident, a software error meant 8,539 audio recordings were not retained; 20 learners requesting a marking review could not have it conducted using all the original assessment evidence.

The limits of this example matter. The published enforcement notice applies to England and concerns ABRSM’s regulatory breaches. These sources concern evidence management, not proof of an AI-grading failure. They do not prescribe the register below or establish a universal retention period for schools, tutors or workplace learning teams.

Our practical inference is narrower: plan for the moment when someone needs to inspect the assessed work again. Keeping only a score, feedback summary or transcript would not satisfy that objective.

Define one evidence package per submission

For this workflow, an evidence package means the linked records needed to inspect the work and follow the assessment history. Store the original as received, then keep later interpretations and decisions separately.

• Original work: the submitted files, including attachments and any assessed recording.
• Submission context: a unique ID, receipt date and time, and the assignment version.
• Rubric snapshot: the actual criteria and scoring scale used, not just a link to an editable rubric.
• Provisional AI output: the saved score, criterion-level comments and generation time. Record the tool and model identifier if available; otherwise write “not exposed”.
• Teacher decision: the final score or outcome, reviewer, review date and explanation of any correction.

Also preserve the instructions supplied to the AI and identify exactly which files or derived text it received. If only a transcript was submitted to the tool, record that limitation rather than implying that it assessed the recording.

Assessment-evidence register template

Copy these fields into an approved spreadsheet, database or assessment system. Use one master entry per submission and linked history entries for subsequent reviews. “Location” should identify a controlled file or record, not merely name a platform.

Add fields for assignment version, receipt timestamp, AI input location, decision status and review-history location where useful. Keep statuses explicit: awaiting AI output, awaiting teacher review, final decision recorded or under review.

A retention-review date is a prompt to assess continued retention. It is not, by itself, an instruction to delete evidence.

Link records without overwriting evidence

Use an institution-controlled location with separate areas for originals, rubric snapshots, AI outputs, teacher decisions and review history. Repeat the submission ID in filenames or record identifiers. This can be implemented through folders or equivalent records in an approved system.

1. Capture the original. Save the submission as received. Record its filename, format and receipt timestamp. Check that every expected attachment opens or plays.
2. Protect the source copy. Restrict routine editing where the system permits. Put annotations, conversions and redactions in a separate working-copy area.
3. Freeze the assessment context. Save the assignment and rubric versions used. Do not rely on a live document that may change later.
4. Save the AI run. Preserve the supplied instructions, input references and returned output. Label the score provisional.
5. Record human review. Save the final decision and rationale separately, then update the register.
6. Test the links. Open the package using an authorised reviewer account rather than the uploader’s account.

This approach is consistent with the ICO’s emphasis on classifying, indexing and locating records accurately. The particular folder structure and register fields are our recommendations.

If a file moves, update its register location and record the change. Ask your technical owner to confirm permissions, version history, backup coverage and restoration procedures; do not assume that a folder labelled “originals” prevents alteration.

Handle written, audio and video submissions distinctly

Written work: preserve the uploaded document or image before annotation. For browser-based answers, save an approved snapshot of the response at submission time. Keep extracted text and converted PDFs separately, identifying their source.

Audio: retain the submitted recording, not only the transcript. Label transcripts as derived records and identify whether they were machine-generated, checked or corrected. Record what the AI received and what the teacher listened to.

Video: preserve the submitted recording alongside any slides or attachments included in the assessment. Keep compressed copies, captions and extracted audio separate. Check playback during intake and retrieval.

For all formats, document transformations. A useful entry might read: “Transcript version 2, corrected by reviewer on 16 October 2026; source recording unchanged.” If a derivative was used for assessment, preserve that exact version too.

Record corrections and appeals as additional events

Do not replace the saved AI score with the teacher’s final mark. Keep both, making their status unmistakable. Record which criterion changed, what evidence the teacher inspected and why the provisional judgement was accepted or revised.

For a later review or appeal, add an event containing the request date, reviewer, evidence inspected, outcome and explanation. Retain earlier decisions according to the approved schedule. If the learner submits revised work, give it a distinct submission ID linked to the earlier attempt.

The objective is reconstruction: a reviewer should be able to distinguish what was submitted, what the AI returned and what the teacher decided at each stage.

Fictional worked example: a spoken explanation

All identifiers, dates, filenames and scores below are fictional. They illustrate the workflow, not an actual learner, product demonstration or retention requirement.

• Submission ID: DEMO-017; a two-minute explanation received on 15 October 2026.
• Original-file location: approved repository, DEMO-017/originals/explanation.wav.
• Rubric version: explanation-rubric-v3, saved in the package.
• AI output: provisional 6/10; full output and input instructions saved as run-01. The AI received transcript-v1, not the audio.
• Teacher decision: 7/10 after listening to the recording. The teacher corrected a transcript omission and explained the criterion-level change.
• Review date: 16 October 2026.
• Access owner: assessment coordinator.
• Retention-review date: 16 November 2026, a fictional administrative checkpoint—not a recommended retention duration.

The package preserves the original audio, both transcript versions, the initial AI output and the teacher’s rationale. A subsequent reviewer can inspect why the two scores differ.

Assign access and retention ownership

As a workflow recommendation, use non-identifying submission IDs and keep the learner-identity lookup separately controlled. Limit routine access to the roles that need the evidence. Do not place unrestricted sharing links in the register or upload learner work to an unapproved AI service.

Ask the responsible privacy or records lead to approve the retention schedule, including originals, derivatives, outputs, review records and backups. Check applicable jurisdiction, awarding-body requirements, contracts and arrangements for unresolved reviews before implementing disposal.

For UK organisations, the ICO’s storage-limitation guidance recommends justified retention periods and periodic review. The page also states that its guidance is under review following legislative changes. Do not treat it as a universal timetable or assume UK requirements apply elsewhere.

Run a retrieval drill before scaling

Give an authorised colleague a submission ID without coaching them through the storage layout. Ask them to complete this checklist and record failures.

• Locate and open every original file.
• Find the assignment and rubric versions used.
• Identify the exact AI input and distinguish derivatives from originals.
• Retrieve the saved provisional output without generating a replacement.
• Find the teacher’s final decision, review date and rationale.
• Follow any corrections or appeals in chronological order.
• Identify the access owner and retention-review date.
• Explain the decision using stored records, without asking the original assessor to fill gaps.

Record broken links, permission problems, missing versions and unclear status labels. Assign an owner to each correction and repeat the drill. Include restoration testing with the technical owner where appropriate.

Keep preparation automation separate from evidence assurance

SubSchool’s AI homework generator creates editable exercise sets from saved lesson context. Teachers can review and change the generated tasks, retaining authorship and the final educational decision.

If repetitive homework preparation is your next bottleneck, explore that feature while implementing this evidence workflow in an approved system. Do not assume the homework generator supplies unchanged-original storage, an assessment register or the retrieval controls described here.