Scarcity is the design constraint nobody names

The argument for building support with learners rather than about them has been settled in the literature for years. What was never solved is the cost.

Doing it properly means sitting with someone across several unhurried conversations, listening carefully, and writing up what you heard without flattening it. Then noticing six months later that it has stopped being true, and doing it again. That is expensive in the one resource the system has least of. Where budgets are tight — which is everywhere — it is the first thing cut.

So the practice became a form. Forms are cheap. They are also the reason a profile can be completed in full and still contain nothing the learner would recognise as themselves.

When software arrived, it optimised the adult’s side of the ledger. Drafting assistants help teachers write goals faster. Document managers help parents track deadlines. Screening aids help clinicians reach a determination sooner.

All three are real improvements to real problems. None of them changes who the work is for, and none of them reaches a family with no institution attached to them in the first place.

The result is a subtler version of the original issue. Authority did move — out of the clinic, toward the family. It stopped there. A parent describing their child is not the same as the child describing themselves, and treating the two as interchangeable is how a field can spend a decade championing learner voice while the learner stays quiet.

Watch what happens around a learner and most of it is procedural: gathering an account of how they think, turning it into a usable document, noticing when it goes stale, revising it, translating it, explaining it to the next adult who needs to understand. Skilled work, but bounded — and scarce because it consumes expert hours, not because it demands an irreplaceable human.

That is the half we build.

The other half is not automatable and should not be attempted: relationship, judgement, teaching, advocacy, and the decisions that have a child’s life in them. When the procedural half stops consuming the week, that is where the hours go. A specialist’s reach multiplies. The specialist does not disappear.

Any system that ends up substituting its own authority for the expert’s has simply installed a new gatekeeper, which is the failure mode we care most about avoiding.

This one is easy to claim and hard to mean, so here is the specific thing that changed our view.

While recruiting for our stakeholder research, one community was almost entirely absent from the responses — not underrepresented, effectively missing. Translating the materials did not fix it, because translation was not the problem.

Two mechanisms were at work. Some families had been advised by clinicians to stop speaking their home language to their autistic child — advice that runs against the evidence, but which has been given for decades and is still followed. Those young people had been moved into English by clinical instruction rather than family choice, so a survey in the heritage language would never reach them. And separately, caregivers in those communities showed the strongest instinct in our entire dataset toward controlling who sees information about their child, alongside the most caution about AI.

That second pattern is not distrust of research. It is a well-described protective norm — a family deciding that an outside institution earns access to a child through a trusted relationship first. Read as a barrier, it looks like refusal. Read correctly, it is a design requirement: build the trust path before asking for the child.

No amount of translated interface addresses either mechanism. That is what we mean by native rather than translated, and it is why cultural context is something the system has to understand rather than something it renders.

This work involves people describing difficult things — sometimes the most difficult things — often young, often without another adult in the room at that moment.

Any system doing this carries a duty that has to be designed for from the start, tested adversarially, and demonstrated rather than asserted. It is also the area where a claim made too early does the most damage, so we would rather be slow and checkable here than trusted before we have earned it.

It is a reason to move carefully. It is not a reason to leave people with nothing, which is the status quo we are measuring ourselves against.

The families who need this most can pay for it least. That is not an unfortunate detail to be solved after product-market fit — it decides what can be built.

A system whose economics require a district budget, a statutory framework, or a family with means will only ever reach people who are already reached. So the target is free at the point of use, funded upstream by the institutions whose mandate this already is: NGOs, agencies and governments.

That only works if the marginal cost is genuinely near zero rather than subsidised down. Which is why we build at the application layer — not training foundation models, not building infrastructure. The useful and unclaimed work is at the top of the stack: conversation design, evidence discipline, document generation, cultural and linguistic fit. It is also the layer where cost can be driven low enough for any of this to be true.

All of the above is a thesis, and a large one. The first test of it is small on purpose.

Sol is being studied with autistic adolescents and their caregivers as doctoral work at the University of Cambridge, under its own ethical approval. Edra is a separate company; the University is not a partner in it.

We will publish what we find, including the parts that do not support the argument above.