Synthetic end to end
The harness never connects to a real school system. Every student, family and member of staff your agent meets is synthetic.
AI test harness · For companies building AI for schools
Put your AI to work in a school with no real students.
Your agent joins a synthetic school as a member of staff and acts through the same channels that role would use. Every action is logged with its effect, and the run report shows what your agent did and what followed.
Status
The harness is specified and comes in the fourth build stage, in the order the first buyers ask for.
Vendors who talk to us now help set that order. Harness runs will be priced after the first pilots.
The interface
Four calls, over REST or MCP.
The interface comes two ways: a REST API, and an MCP server so any agent framework can connect without custom code. Both use one message schema.
Illustration · one morning in a harness run
- Your agentRegisters as the attendance officer.
- The schoolSends the morning's register export. Two marks are missing and one code is wrong, as registers sometimes are.
- Your agentMessages a family about an unexplained absence.
- The engineResolves the message by rule. The family's reply arrives later, through the parents' channel.
- Your agentCalls for the report: every action, its effect, and whether each procedure step was followed.
The run report
See what your agent did, and what followed.
- Every action your agent took, logged with its effect in the engine.
- Outcomes for the run, and whether each procedure was followed.
- Reviewers can flag any message your agent sent.
- An audit log records every model call and admin action.
Why a synthetic school
A school's real pressures, before any real student meets your product.
Messy, the way schools are
Your agent meets late registers, miscoded records, upset families and procedure steps it must not skip, all drawn from the pack for that school system.
Repeatable
Every run replays exactly from its seed and can branch from any save point, so two versions of your agent face the same engine draws. Where their actions differ, the cast replies afresh.
Harness tenants authenticate with scoped keys and are rate-limited, and row-level security keeps each tenant's runs apart.
The rules your product meets
Rules for AI in schools are arriving, and some ask for testing.
| Where | What it asks |
|---|---|
| England | The Department for Education's product safety standards, which are non-statutory guidance, ask for testing "with a diverse and realistic range of potential users and use cases" (DfE). |
| European Union | High-risk systems, which include four education uses, must be tested against defined metrics before they reach the market (AI Act, Article 9). Those obligations apply from 2 December 2027 (Lewis Silkin). |
| United States | The Department of Education left safety rules for AI in schools to states and districts in its 2026 grant priority (Federal Register). Ohio, for one, required every school district to adopt an AI policy by 1 July 2026 (Ohio Revised Code 3301.24). |
| Abu Dhabi, UAE | ADEK now requires private schools' AI learning tools to meet its requirements (Abu Dhabi Media Office). |
| Elsewhere | Other education regulators are still writing their rules. Tell us which system your product is for, and its rules shape what the harness tests first. |
| China | A 2025 guide asks education departments to keep a whitelist of generative AI tools allowed in schools (eol.cn). |
A harness run gives you evidence for your own testing. It does not certify a product or show that it complies with any law.
Shape the harness with us.
Tell us what your product does in schools and which school system it is for. The first vendors' needs set what the harness does first.