Audit equity and access after the project

1

Research

Why this matters, and what good looks like

An equity audit asks whether the project's arrangements created barriers to participation or access. This method examines one material and one barrier after a project. EEF provides evaluator guidance, Project Zero a content-representation routine, and DfE England-specific data-protection guidance. These sources do not validate this combined equity audit or establish AI benefit. They provide no universal safe small-group threshold. Even aggregate information can identify people.

EEF includes reach and responsiveness among implementation dimensions. This gives no permission to analyse identifiable groups.
Project Zero asks whose perspectives are missing and encourages speaking from one's own perspective.
DfE calls for transparent handling and local data-protection or IT guidance.

The stricter minimisation gate below is a local process choice. Actual affected voices and co-design remain outside AI.

Where the evidence comes from

Moderate-strong for inclusive improvement; AI benefit unvalidated and high-risk.

2

Workflow

Brief it, steer it, check it

Apply the private ethical gate

Before AI, define purpose and safeguards with local responsible staff. Check one material and minimise evidence privately. Exclude individual records, small groups, identifying aggregates and proxies, meaning indirect identity clues. Unknown safety means stop before uploading.

Prompt 1 · Gate

CONTEXT
Non-sensitive material excerpts with IDs and versions: [material].
Human ethical-gate record, containing only purpose, permitted material, minimisation outcome and local review status: [gate].
REQUEST
Check whether the supplied record permits this bounded material-only comparison. List missing safeguards without requesting sensitive detail.
QUALITY BAR
No individual, small-cell, proxy or protected-trait data or inference. Consent alone is insufficient; aggregates may identify. Never infer safety or certify compliance. If approval, minimisation or local review is missing or unknown, return HOLD and stop analysis. Do not repeat unsafe details. Describe only supplied human decisions.
FORMAT
Table: Gate item | Record ID | Supplied decision | Status | Human action.
Cover Purpose, Minimisation, Local review. Status is Supplied only or HOLD. End with Scope limit.

Examine the material barrier

Check the gate output privately. Optional access plans and evidence reviews can help; otherwise use your own checked project material and safe process observations. Reapply minimisation even to prior outputs. Keep all learner-level evidence outside AI.

Prompt 2 · Barrier

CONTEXT
Exact approved gate output: [gate output].
Versioned teacher gate corrections, safe material excerpts and permitted non-identifying process evidence: [evidence].
REQUEST
Identify one evidence-linked material barrier and offer two possible responses for human consideration. Separate observed arrangements from possible effects.
QUALITY BAR
No attribution of ability, protected traits, identity or deficits. No individual or small-cell analysis, proxy inference, grading or causal conclusion. Keep dissent and missing evidence explicit. Missing gate approval means HOLD only. Do not invent affected voices or choose a response.
FORMAT
Table: Barrier ID | Exact material quote | Evidence ID | Observed arrangement | Possible effect and limit | Response options | Missing voices or evidence.
Use B1. End with Human review required.

Co-design outside AI

Meet affected learners through an accessible, appropriate local process. Check the barrier and co-design the response, retaining disagreement. Return only a safe decision summary. If voices are absent, stop co-design claims and keep the response on HOLD.

Prompt 3 · Record

CONTEXT
Exact approved barrier output: [barrier].
Versioned teacher corrections and safe summary of actual affected-learner review and co-design, chosen response, unresolved dissent, owner role and monitoring measure: [co-design].
REQUEST
Record the human-selected response in a bounded equity audit, separating participation evidence limits, material evidence, affected-voice process and next monitoring.
QUALITY BAR
Do not impersonate learners or manufacture consensus. Do not accept AI role-play as affected voices. No personal data, identifiable aggregate, proxy or protected-trait inference. Missing actual affected participation, selection, owner or measure means HOLD. Label supplied decisions as supplied, not independently verified. Do not promise equal access or improved outcomes.
FORMAT
Table: Barrier ID | Participation evidence and limits | Material evidence ID | Affected-voice process and dissent | Selected response | Owner role | Monitoring measure | Status.
Status is Human-selected proposal or HOLD. End with Unknowns.

Check the selected response

Privately check the audit against original records with affected learners and responsible staff. Submit only safe edits for comparison. Humans authorise any trial. If the bounded work exceeds 60 active minutes, stop and arrange further participation rather than bypassing it.

Prompt 4 · Audit

CONTEXT
Exact teacher-selected audit and versioned safe edits: [audit].
Safe source excerpts and human verification summary covering minimisation, affected voices, dissent and monitoring: [verification].
REQUEST
Compare this selected response with supplied evidence and human decisions. Flag exclusions, invented consensus and unresolved privacy or access issues without rewriting it.
QUALITY BAR
No identity, proxy, protected-trait or deficit inference. Unknown safety or absent affected voices means HOLD. Never treat an aggregate as automatically safe or a teacher summary as independent proof. No legal certification, AI grading, causal impact claim or release approval. Do not repeat unsafe text.
FORMAT
Table: Check | Exact safe audit quote | Evidence ID | Finding | Human action.
Cover Privacy, Evidence, Affected voices, Dissent, Response, Monitoring. End with Unresolved holds and Human release remains required.

Check before you use it

Did responsible humans confirm minimisation and local safeguards before any material entered AI?
Could any aggregate, rare combination or indirect clue identify someone? If unknown, keep it outside AI.
Are findings about arrangements rather than inferred traits, ability or deficits?
Did actual affected learners review and co-design outside AI through an accessible process?
Are disagreement, missing voices and evidence limits retained without invented consensus?
Have humans checked the selected response, owner and monitoring measure before authorising a trial?

Stop rule

Stop if small-cell data risk identification, proxies encode bias or recommendations exclude affected voices.

3

Reuse it

So the next one takes minutes

Not suitable

Build a skill

Save this as a reusable instruction you can run on any project.

Sensitive context and actual affected participation require a fresh human gate each time.