Haven

Haven is a companion for the waiting period. A young person describes what is going on, and Haven points them to a specific named service they can use tonight, such as NHS Talking Therapies, Kooth or Shout, or back to their GP where the referral needs revisiting. It does not treat, and it does not pretend the wait is not happening.

The product is a small illustrated village with five places in it, not a chat window. The target user has already decided not to phone anyone, so a clinical form loses them at the first field.

Only one of the five places calls a model at all. The Lighthouse, the Library, the Signpost and the Common Room are fixed content: crisis numbers, breathing exercises, a service directory, other people's accounts of waiting. That ratio was the design, not a scoping compromise. Almost everything a person needs during the wait is retrieval, and retrieval does not need a model.

Inside the Listening Post there are three model calls, and only one writes anything the user reads as conversation. The other two are a one-word safety classifier and a bounded reformatting of what the user already said.

Breathing exercises and a service directory live elsewhere in the village, with no model in either. That is what four of the five places look like. The Listening Post is the exception, and even there most of the work is constraint rather than generation.

Risk language is detected before the completion call and short-circuits it entirely. The user lands on a fixed screen with Samaritans, Shout and 999. No generated text is involved in the highest-stakes moment in the product. A model that is right 99 times out of 100 here is not good enough, and it cannot be made good enough by prompting.

Not because the number is optimal, but because an uncapped supportive conversation with a distressed person becomes therapy by drift. The cap forces a routing decision while the exchange is still small. Three turns of listening, then a routing decision.

Every exit terminates at a human service. Haven never resolves a conversation itself. The output of a successful session is a named organisation, not advice.

The routing logic distinguishes "this has lasted months and is affecting your sleep, your schoolwork and your friendships" from "this is a hard fortnight", and when it routes to a GP it produces a list of things to say, drawn from what the user told it. Duration, sleep, "I stopped going to lessons." The list is model-written and bounded to what was said. That last point renders as "The impact on your studies." It never names a condition, because it is a description of the conversation, not a conclusion about the person.

The waiting list is under-18s, which means a safeguarding protocol is not a later refinement, it is a precondition for a single real user. Haven does not have one. Building for this group in a day meant building something that cannot be deployed to the group it is for, and that constraint shaped what went in more than any other.

And one that is not a constraint. The Common Room is fixed content, shared accounts of the wait, no generation involved. It exists because the loneliest thing about a waiting list is assuming you are the only one on it, and that is a product problem rather than a clinical one.

A six-hour prototype with a commissioner dashboard running on mock figures.

The metrics on it are the right ones to be judged on, and none of them have been measured. Drop-off and time to first contact are the two Haven could not measure even with real users, because it has no visibility into the NHS side of the referral. It knows who opened it. It cannot know who eventually got seen. A commissioner view that reports on outcomes it structurally cannot observe is the exact failure the rest of this project is about, and it is sitting in the build.

Nothing here has been near a patient. Beyond the safeguarding protocol, it would need clinical governance and an evaluation of the routing logic against clinician judgement before it went anywhere.