← Industries/EdTech
Promethean for EdTech
AI Act Annex III §3 makes education-AI high-risk: access determination, assessment scoring, monitoring of behaviour during exams. GDPR Art. 8 + 22 layer on top for under-16 users. FERPA in the US adds another evidentiary surface. Promethean is the per-decision audit layer that makes 'we know every grade or recommendation the AI produced for every student, and a human reviewed the appeals' structurally true.
Who this is for
Compliance + product teams in edtech SMBs (adaptive learning · assessment AI · proctoring · tutoring). Typically:
- Seed–Series B edtech with 5–60 engineers; 1–4 LLM features in production.
- Customer base mixes K-12, higher ed, and corporate L&D — each with different audit cadence.
- Per-student volume: 10–500 AI-touched decisions per student per month (grades, recommendations, flags).
- Multi-tenant by institution; multi-tenant within institution by class / cohort.
- FERPA / GDPR Art. 8 require special handling for under-16 users + parental-consent paths.
The regulatory pressure
AI Act Annex III §3 — education + vocational training
AI for access-determination to educational institutions, assessment scoring (formative + summative), monitoring of behaviour during tests is high-risk. Full Chapter II obligations apply.
Deep-dive →GDPR Art. 8 + 22 — child-data + automated decisions
Under-16 users (or younger per member-state lawful age) need parental consent for processing personal data. Grading or admissions decisions that produce legal effects fall under Art. 22.
Deep-dive →FERPA (US) — Family Educational Rights and Privacy Act
Educational records protected; parental/student access rights to records the AI generated. Per-decision logs underpin the access-rights workflow.
COPPA (US) — children under 13
Children's Online Privacy Protection Act mandates parental consent + data-minimisation for under-13 users. The substrate's PII-safe-by-construction design helps demonstrate min-necessary.
Where LLMs typically run in EdTech
Automated essay grading + feedback
Classifier producing rubric-based score + comment per essay. Always-human reviewer gate on grade appeals; the rubric (spec) is hash-committed so silent rubric changes are detectable.
Adaptive next-question routing
Router selecting next problem from a difficulty + topic pool given the student's history. Schema-bound output; chain shows fairness across student cohorts.
Plagiarism / academic-integrity flag
Classifier flagging suspect submissions. Always-human reviewer gate (faculty review every flag); reviewer-verdict captured for appeals.
Tutoring chatbot
Drafter producing tutoring responses. Schema constraint on age-appropriate language; reviewer-gate on flagged outputs; fallback to canned safe response on schema failure.
How the substrate maps to your audit
| Your LLM feature | What the regulator asks | Promethean evidence |
|---|---|---|
| Essay grading | Per-essay evidence of grading consistency + reviewer involvement on appeals. | specHash binds the rubric; reviewerVerdict captures appeals; chain shows grade distribution. |
| Adaptive routing | Fairness: are students from different cohorts getting comparable difficulty? | tenantId per student/cohort; statistical analysis against chain. |
| Plagiarism flag | Faculty review evidence per flagged submission. | Always-human reviewer gate; reviewer identifier captured; appeals trail preserved. |
| GDPR Art. 22 appeals route | Meaningful human intervention available when the student contests. | Reviewer-gate firings + verdicts; override rate computable from chain. |
| FERPA access rights | Student requests records of AI-influenced decisions on their record. | Filter chain by tenantId + recordedAtIso; export per-student decision metadata. |
Which Promethean tier fits
Recommended for typical SMBs in EdTech
Team€79 / spec / month
Unlimited specs; €79 per active spec per month. 100k entries per spec. Daily OTS anchoring. Single-tenant.
Most edtech SMBs start with 1–3 specs (grading + routing + tutor) — Team tier at €79/spec/month covers this naturally. Upgrade to Production (€499 flat) once you cross ~6 active specs or need multi-tenant primitives for institution-level isolation. Enterprise only matters once you're selling into government education ministries needing the framework-handoff packets.
Larger EdTech operators with multi-tenant or framework-template needs upgrade to Production (€499 / month flat).
What this looks like in practice
Hypothetical: an adaptive-learning SMB during a school-district procurement
A K-12 adaptive-learning platform competes for a 200-school district contract. The district's procurement RFP includes: 'demonstrate AI grading is consistent across demographic groups; demonstrate appeal process is meaningful; demonstrate FERPA-compliant audit trail per student.' Without Promethean: written assertions + a one-time fairness study. With Promethean: an audit bundle showing per-student decision counts, override rates on appeal, fairness statistics computed against the chain. The district's IT-procurement officer runs verify.mjs against a sample. The conversation moves from compliance-by-promise to compliance-by-evidence; the SMB wins the contract.
Frequently asked
We serve students under 16. Is Promethean appropriate?
Yes — the substrate's PII-safe-by-construction design fits child-data protection well. Raw inputs (student work, names, etc.) never enter the L12 chain; only SHA-256 hashes. GDPR Art. 8 (child data) and COPPA (under-13 in the US) both favor data-minimisation; Promethean structurally enforces it. Your platform still handles parental consent + data lifecycle on the source data; Promethean handles the audit-trail layer.
Schools want SOC 2 + state-level edtech privacy assessments. How does Promethean help?
Promethean's chain provides per-decision evidence that flows into state assessments (e.g. SOPIPA in California, Texas SB 820). The audit-bundle download is the artifact most state edtech privacy assessments ask for — proof that AI-driven decisions are traceable + reviewable + bounded. Doesn't replace SOC 2 (organisational controls), but materially reduces the AI-specific evidence burden.
What about AI-generated tutor responses — those aren't 'decisions', are they?
Tutor outputs are still decisions under AI Act Art. 50 (transparency obligations for AI-generated content interacting with humans). Article 50 requires disclosing to the user that they're interacting with an AI system. The L12 chain captures every tutor response with the spec + model identity + schema-valid flag — so when a parent or regulator asks 'what did the tutor say to my child on date X', you can produce the exact record (input/output hashes link to your retained source data).
We use a third-party LLM provider. Does that create extra audit complexity?
It creates third-party-risk evidence — but Promethean addresses it. The modelIdentity field in every L12 entry records the exact provider + model + version per decision. Procurement officers can verify your vendor relationships at the per-call level. DORA's Art. 28 third-party-risk requirement (for fintech) has parallel obligations in many edtech state privacy laws; the chain answers them.
Do we need separate Promethean workspaces per school district customer?
No — one workspace, multi-tenant. The R4 tenantId field on each L12 entry identifies the school district (or class, or cohort — whatever your tenant model is). Production tier and above support this natively. Each district can be issued an audit bundle filtered to their tenantId; they verify with verify.mjs without seeing other districts' data.
Definitions used on this page
The substrate primitives referenced above (L12 receipt chain, spec hash, reviewer gate, fallback behaviour, OpenTimestamps anchor, tenant ID) all have canonical definitions in the glossary: