The deadline was doing the work.

Five things landed in front of English schools in the last week of September 2026. Separately they look like unrelated admin. Together they are one argument about what keeps a record of a child honest.

Start with the smallest. The Department for Education's update of 23 September reminded schools that the PE and sport premium expenditure return was due at 5pm on 30 September. It asks how the money was spent and what impact it had. A school answers the first from invoices in a few minutes. Answering the second means knowing who actually took part, broken down so pupils with special educational needs are visible. Not who was eligible. Not who was invited. Who went.

Then the inspection toolkit. Ofsted's schools inspection toolkit, version 2.0, has been in force on inspections since 7 September 2026. Its inclusion evaluation area asks a school to show that pupils with barriers to their learning or wellbeing are achieving well, and it refers to leaders who have an inclusion role rather than the SENCO alone. That is right in principle, and it distributes the evidence across people who have never had to assemble it into one answer. Ten people each holding part of the picture is not a picture.

Then the consultation on education otherwise than at school, closed on 18 September 2026 with a response expected this autumn. It asked who should commission and quality assure provision for a child whose needs cannot be met in a school. Whichever way it lands, the school holds the relationship, the local authority holds the arrangement, and the complete record sits in neither place.

Then the Information Commissioner's Office guidance on how the Children's code applies to education technology. It is written for suppliers, but read from the buyer's side it becomes a list of questions a school may ask before signing: controller or processor and on which processing, what the contract says about retention, and what happens to the data when the contract ends. The audit findings behind it, published on 24 June 2026, came from 28 providers, so none of it is hypothetical.

And then the one that changes the shape of all the others. The education for all bill, announced on 13 May 2026 and reported by Schools Week, will require an individual support plan for every child identified with special educational needs, introduce a national template for education, health and care plans, and replace the annual review with a review at the child's next transition point.

That last change is presented as a workload saving, and narrowly it is one. But an annual review is not only paperwork. It is an audit. It forces a record to be brought up to date once a year whether anybody felt like it or not. Remove it and a plan can run from the end of one key stage to the end of the next with nobody obliged to look at it.

So the week pulled in two directions at once. One strand of policy is removing the deadline that produced a current record. Another strand, inspection and the funding returns, is raising how often somebody has to produce one on demand, and widening the group of people expected to hold it. They are the same unresolved disagreement: is evidence something you assemble when asked, or something simply true at any moment because it was recorded as the work happened?

What we are working on

Aevum Foundry AI is in development. What follows is the problem we are working on and how we are thinking about it, not a finished product.

Almost everything a school knows about a child with special educational needs is already written down somewhere. It is in a club register, a behaviour log, a note from a teaching assistant, the paragraph a SENCO wrote in March. None of it is wrong. It is simply not in a form anybody can ask a question of, which is why assembling an inclusion picture costs a day a term, and why it is only ever as current as the last deadline.

Two constraints shape how we are approaching that. The first is that the sensitive paragraphs must not leave the building. A behaviour log or a support plan is special category data about a named child, and the ICO's audit findings describe what happens when that depends on somebody else's contract. So the processing runs on hardware the school controls. The second is that a person must be able to see what produced any output and overrule it, because a record that is continuously correct is worth nothing if nobody can tell where a line came from.

We are not claiming local processing is sufficient. A systematic review published on 28 May 2026 found the literature on attacks against on-device AI badly out of balance: around half of all defence work addresses model theft, while adversarial attacks, the ones that produce a wrong answer, have almost no matching defence work at all. Running on your own hardware answers where the data went, not whether the answer is right. Both questions have a child at the end of them.

So the honest version of what we are aiming at is this. If the forcing function goes, something else has to keep the record honest, and we would rather that something was the work itself than another deadline.