Guide · DfE standards updated 19 January 2026 · 13 areas, 87 requirement rows

Evidencing the DfE generative AI product safety standards

Name: DfE AI safety evidence toolkit
Price: 149 USDC

The DfE expects suppliers of generative-AI products for schools to show, with evidence, how they meet its product safety standards. This toolkit turns the standards into an evidence matrix, risk register, DPIA input pack, release tests and a procurement summary you fill in yourself.

Not certification. Not certification, accreditation, legal advice, a DPIA, data protection officer sign-off, safeguarding assurance or a safety guarantee, and not endorsed by the DfE. Pupil-facing products should also get qualified data protection, safeguarding and child-development review.

What the standards ask for

The DfE Generative AI: product safety standards (first published 22 January 2025, updated 19 January 2026) cover 13 areas and put the burden of evidence on suppliers: DPIAs, release testing, transparent claims, security, monitoring, intellectual property and attention to cognitive, social and mental-health harms.

Stated purpose
Educational use cases
Filtering
Monitoring and reporting
Security
Privacy and data protection
Intellectual property
Design and testing
Governance
Cognitive development
Emotional and social development
Mental health
Manipulation

Why evidence matters now: the ICO's review of 28 widely used edtech providers (ICO, 24 June 2026; key findings PDF) made 596 recommendations and found that almost 90% had incomplete records of processing. DfE's standards page remains the authority; this toolkit paraphrases it under the OGL.

What is in the toolkit

Intake questionnaire (written, 8 sections, document checklist)
Evidence matrix: 87 requirement rows across the 13 areas, each with the evidence a supplier would normally show, status, gap, action and owner
Risk register with scored ratings and 15 common generative-AI school risks, plus a complaints log
DPIA input pack following the ICO steps (input to your DPIA, not a DPIA)
Data-flow register and diagram template
Release test protocol and a 68-test release test set (filtering incl. misspellings and other languages, disclosures, jailbreaks and prompt injection, privacy canaries, IP, accuracy and bias, cognitive offloading, dependence, distress pathways, manipulation) with go/no-go rules
Complaints and escalation workflow; procurement summary template
The worked sample below, a licence and a SHA-256 manifest

Free worked sample: our own product, gaps included

The sample applies the toolkit to Adam Hale's Report Writer. Across 87 rows: 11 met, 20 partly met, 7 gaps, 1 evidence needed, 48 not applicable. It includes a lexicon scan of 39,611 sentences and 7 of 7 live browser tests passing, and lists the gaps honestly (for example, no supplier DPIA yet).

Download the worked sample (ZIP)

Buy the toolkit, 149 USDC

DfE AI safety evidence toolkit

149 USDC on Base

File: DfE-AI-Safety-Evidence-Toolkit-USDC.zip, 2.1 MB. SHA-256 865303f832efc795a747a400373bd927c5558af06f2dfbd40c3fa869d9009fc5

Option 1: plain USDC transfer (any wallet)

You get an order page with an order-specific amount (149 USDC plus at most 0.009999 USDC, so the payment identifies your order), valid 24 hours. Send it as one canonical USDC transfer on Base, paste the transaction hash, and the download appears after 10 confirmations.

Option 2: x402 (agents and x402 clients)

GET https://pay.adamhaleworks.workers.dev/x402/v1/dfe-ai-safety-toolkit

Returns HTTP 402 with x402 v2 "exact" terms (149000000 atomic USDC to 0x491Fcf30a5aa0483B3cd8C0041d21801F195B3AF). The official x402 client caps automatic payments at $1 by default; raise it for this purchase, for example x402Client.fromConfig({ …, spendControls: { maxAmountPerPayment: "$150" } }), or use option 1.

Licence: /licence/ai-safety-toolkit/. Download links last 48 hours and are re-issued from the order page for 30 days. Payment goes straight to the seller's own wallet. There is no merchant of record, card processor or platform in between, and payments on Base are irreversible.

Need help filling it in? The written diagnostic and full evidence pack are offered on request at ai-safety-evidence.adamhaleworks.workers.dev; they are not sold here.

Attribution

Contains public sector information licensed under the Open Government Licence v3.0. (DfE, Generative AI: product safety standards, updated 19 January 2026; OGL v3.0). Built and operated by AI agents working for Adam Hale. Data is compiled by code from official sources and checked by scripts; no person reviews individual answers or files.