The evidence
What helps a child learn.
And what keeps them safe.
The research and the principles behind Aevum: sovereign, local-first AI built for classrooms and for the children the average tool leaves out.
The problem
The most careful place to draft a child's plan is the room it is about.
Drafting a support plan means handling exactly the information a school least wants to send anywhere: a named child, their needs, the support they receive. New Department for Education guidance tells schools they may use AI for this work, and in the same breath to be extremely cautious about putting that child's data into it. Both things are right. Together they are the design problem.
1.7m
Pupils with SEN in England (DfE, 2024/25)
Why local-first
The safest model is the one already running in the building.
Aevum is built local-first: a child's data never has to leave the building to be useful, and it never comes to us. When the data does not travel, the warning to "be careful" takes care of itself, and the school keeps control of what it is responsible for.
Sovereign
Designed in the UK · Trained on UK national compute
Faithful, inspectable reasoning
An answer a teacher can't inspect is an answer a teacher can't use.
In a classroom, trust is not a feeling, it is a requirement. A teacher needs to see why a tool reached its answer, to check it against the sources that are allowed, and to know when it is unsure. Aevum is built so its reasoning can be inspected and its confidence can be acted on, because in education an unexaminable answer is not good enough.
Inspectable
Reasoning a teacher can check
How this shows up in the classroom: The Teaching Manager →
SEND-first by design
Accessibility is the architecture, not an afterthought.
Most technology is designed for the average user and adapted for everyone else afterwards. Aevum starts the other way round. It is designed with disabled, SEND and AAC users as the first users, not a later accommodation, because a tool that works for the child at the edge works for everyone. Nothing about them, without them.
51 hrs
Average teacher working week; ~21 hrs actually teaching (DfE Working Lives)
How we prove it
We publish the test before the result.
Aevum makes strong claims about faithfulness, calibration and safety, so we hold ourselves to evidence. Our formal evaluations are underway on UK national compute, and we register what we are going to measure before we measure it. When the results exist and stand up, they will appear here, in full, including where they fall short. Until then, the honest status is: designed to, and being tested.
In progress
Evaluations underway · Results published when they exist
The standard we set ourselves
We publish what we can prove, and we say plainly what we can't yet.
One email, with the link, when each thing is ready.
Last reviewed
Changelog: first publication of this page (25 August 2026). No results published yet, by design.