Moorwell does not write to you. It reads what you wrote, works out what it is allowed to offer, and picks one. Every sentence it can say already existed before you opened it, and a person could read all of them.
The stages run in this order every time, on every input. Nothing later can undo something earlier — a claim you could check by reading the code, not a diagram drawn after the fact.
A word, a sentence, a paragraph. It is read and answered on the device. What you typed picks a reply and is then gone — not written to a database, not synced.
Held back: nothing. The person is the only source of input — there is no sensor, no microphone and no background reading of anything.
The risk check runs first, on every input, always. It is not one option among several planners, it cannot be switched off, and it cannot be swapped out. It is a table of reviewed phrases matched as whole words, with negation handled — deliberately a list rather than a model, so a clinician can read it and say exactly what it will and will not catch.
Held back: the phrases, and every rule about them. What is publishable is the shape and the honesty about its limit: the project writes this down as a floor, not a guarantee, and the always-visible route to help exists precisely because this layer will sometimes miss.
Not a diagnosis and not a score. A band is a coarse reading of how much room there is for anything demanding, and it decides what the libraries are willing to hand over. Where a model has a say at all, it is allowed to move the reading in exactly one direction: it may raise an assessed level of distress, and it may never lower one.
Held back: the bands, their boundaries, and what moves a reading between them. The one-way rule is the part worth publishing, because it is the part that is checkable: it is written once so no call site has to remember it, restated at six more places in the code as defence in depth, and held by its own test. A model that could lower a reading turns a misread into a withheld crisis resource, and that is the one failure this app cannot absorb.
Hard rules live here and nowhere else. If the answer to may this ever appear in this state? is no, the item is filtered out — never scored down. The gates sit inside the libraries themselves rather than on the screens that call them, so a screen cannot get it wrong by forgetting.
Held back: which rule applies where. The reason for it is a bug this project shipped: a hard rule expressed as a score penalty is not a hard rule, because under any ranking penalised means unlikely rather than never — and a game once appeared on a good day for exactly that reason.
Only preferences live here — what tends to suit you, what you said you care about, what fits the length of time you have. Ranking sorts inside the gate and can never open one: a better-matching line at the wrong band loses to a worse-matching line at the right one.
Held back: the scoring. The output type is the most important fact on this page: the last stage returns an id, not a sentence, the identifier of an entry in the bank. That is what makes every word the app can say reviewable in advance.
Each is an absence in the build, which is a stronger promise than a policy.
Nothing is composed at runtime. Every string ships in the app and can be read before you install it.
Assessment runs on the device. Aeroplane mode changes nothing about the reply.
A band is not a condition. Moorwell never names one, and never implies you have one.
A model in this system never produces the words you read. It is handed a slate of things it may choose from, and what it returns is a choice — validated against the catalogue before anything is drawn, and discarded whole if it does not match.
"You've had a hard week. Try being kinder to yourself tonight."
Prose, composed on the spot, arriving at a person who may be having the worst evening of their year. Nobody reviewed it, because it did not exist until the moment it was shown.
{ "pick": "val-x4" }
An identifier from a slate the engine built. The entry it names was written by a person, carries its own review status, and sits behind the same gates as everything else. If the identifier is not on the slate, the whole response is thrown away rather than partly applied.
As built today there is no wire field that could carry prose in the first place — the type has nowhere to put a sentence. The validator exists anyway, and the on-device and cloud paths go through the same one, so the remote path can never be checked less carefully than the local one.
Everything it can say ships inside it, which is what makes reading all of it possible.
A clinician can read every sentence this app is able to say.
That is a property no system that writes its own words can offer, at any budget, because the sentences do not exist until they are shown.
The packet a clinician would read is built and current: the 2,227 entries counted above, sorted into 232 groups, every one gated and traceable. What that count does not tell you is who did the checking.
The risk check is deliberately rule-based, and recall is chosen over precision.
A false positive shows somebody a route to help they did not need, which is mildly patronising. A false negative misses somebody who did. Those are not symmetrical, so the design is not symmetrical either.
The code says so in its own header, and it gives three reasons. The first is auditability: a neural decision boundary cannot be reviewed the way a table can, and this is the one part of the app where being wrong is not recoverable. The second is the asymmetry above. The third is the limit — keyword matching cannot catch every disclosure, so the project records it as a floor rather than a guarantee.
There is no model anywhere in the risk path. Its only imports are language-pack utilities; there is no embedder, no runtime, no inference of any kind. One diagram in the project's internal documentation draws an optional embedder into that box, and it is wrong — the path does not exist, and the code and the safety-invariant note both say so independently.
Moorwell does not go after a diagnosis. It goes after the loops that keep a bad week going — the ones that quietly make tomorrow harder than it needed to be. Every part of the app names the loop it interrupts.
Withdrawal is the easiest one to see, so start there. Read it round once, then cut it.
Two techniques are not aimed at a loop at all — they are for the ten minutes you are in. Paced breathing slows you to about six breaths a minute. Grounding gives your eyes and your hands a job, so the loop has less room to run. The five-four-three-two-one form of grounding is taught nearly everywhere and has not been tested much on its own.
Each was measured rather than estimated, and each is a claim this site could have made without anybody catching it.
Each choice below cost something that would demo well and bought something better in its place.
The rules are the product and run on their own. The two models sit on top of them to catch what a rule written in advance cannot: the phrasing nobody thought of, and the feeling underneath it.
A deterministic planner over a bank a person wrote — the five stages above. It is the product: the app is complete with both models below switched off, and it behaves the same way twice on the same input, which is what makes it reviewable at all.
Small, open-licensed, and from the retrieval family rather than the generative one. Its job is to notice that what you wrote and something in the bank are about the same thing even when they share no words. It changes which line is chosen — never what the line says.
It reads for feeling, and its promise is precision on the readings it accepts rather than accuracy on all of them — so where it is not confident it abstains, and the engine proceeds without it. It may raise how much difficulty the app assumes you are in. It may never lower it.
Both models run on your device, and neither produces a sentence you read. And neither is anywhere near the part of the app that reads for danger — that path is keyword matching over reviewable phrase lists, with no model of any kind in it, because it is the one place where being wrong is not recoverable.
