What is built, and what is not.

Some of this works today. Some is only designed. Some we genuinely do not know yet. Here is which is which - you should not have to work that out for yourself.

Working now

The clinician's app

Taking notes from a conversation, storing them encrypted on the machine, and running the AI and transcription locally all work today. That is our own assessment; nobody outside has checked it.

Designed, not built

The part that makes us different

The ranking maths, the way promises are described, the safety rules and the testing plan are all written down in full and internally checked. Independent review has not happened yet.

Not built

The ranking engine

The maths that decides what is most urgent does not exist in code yet. Neither does the shared safety core, or the store that holds every promise.

Unproven

Whether it helps patients

We do not yet know if it works. Accuracy, whether it holds up in another hospital, whether it improves care, whether it works equally well for everyone, and what regulators will say are all still open.

Safety rules

eight, non-negotiable

Rules the build has to pass.

These are the rules we hold ourselves to, not a claim that we already meet every one. Breaking one should stop the build - not turn up as a reassuring footnote three months later.

01Patient data is never stored unencrypted
02Nothing leaves the machine without consent
03A person signs off on every clinical action
04Every change is recorded and traceable
05When unsure, it stops and asks a human
06Tested for fairness across six groups
07It fails loudly, never silently
08Regulatory controls are designed in from the start

Tests it must pass

prove it first

What has to be true before it goes near a patient.

A model can be accurate and still be useless, or harmful, if it buries the people using it in alerts. Both have to pass.

Area	Gate	Detail
Agreement	Two reviewers agree at least 70% of the time	If experts cannot agree what counts as a miss, nothing else matters
Fit	Predictions line up with what happened	No systematic drift when checked
Calibration	A stated 70% chance means roughly 70%	Within a narrow margin, not loosely
Interruptions	No more than 6 alerts a day	Per coordinator - more than that and people stop reading them
Warning time	At least 24 hours notice, typically	Late enough to be wrong, early enough to act on
Fairness	Works equally well across six groups	Checked separately for each, not just on average

How this could fail

how it dies

Suppose this does not work. What went wrong?

Every one of these has a warning sign we can watch for, and something we would do about it. The point is to stop money going into a broken idea, not to look confident.

PM-01

We cannot agree what counts as a miss

Warning sign: Two experts look at the same case and disagree too often.
What we do about it: Have people judge cases by hand before building anything. If they cannot agree, stop.

PM-02

The patent covers the wrong thing

Warning sign: Our lawyer cannot match what we have built to what we filed.
What we do about it: Sort out what the patent actually covers before showing the details more widely.

PM-03

Old problems eat the roadmap

Warning sign: Known safety issues stay flat or grow while new features keep shipping.
What we do about it: Make fixing them a condition of releasing anything, and check it more than one way.

PM-04

It gets regulated as a medical device

Warning sign: Our lawyer cannot argue a clinician could independently review what it suggests.
What we do about it: Keep the useful everyday parts separate from the prediction, and plan the regulated route in advance.

PM-05

It nags people until they ignore it

Warning sign: More than six interruptions a day, or people overriding it more often.
What we do about it: Treat alert volume as a reason not to ship, not something to tune afterwards.

PM-06

We never get enough data to prove it

Warning sign: We need more cases than we can get, and no second site has agreed.
What we do about it: Work out how much data is needed early, and fund a second site as a real milestone.

Source notes

checkable

The rules we are building against.

These are the outside rules we are designing against. Our own capability claims come from an internal document and have not been audited by anyone independent.