For GP trainees, GPs, and the people who teach them
Know when you know enough.
The best GPs are not the ones who ask the most. They know which question is worth asking, when the answer has crossed the line, and how to let a patient go safely. That is what Threshold trains. Ten minutes a day.
Do today's ten minutes How we mark
No language model marks you. No checklist either. Every mark shows its working, and you can check it.
Today's call · 2026-09-22
A 24-year-old attends with a three-day sore throat. No cough or coryza. Fever in the last 24 hours. On examination the tonsils are severely inflamed with purulent exudate. No red flags. Rapid antigen testing is not available.
How likely is it a strep throat?
In our showcase world (teaching numbers, declared as such, with a starting likelihood of 20 to 35 percent), these findings push it to roughly 60 to 75 percent. The checked reference gives an antibiotic now; a back-up prescription is the nearest also-sound call.
Teaching numbers from the showcase world, not clinical evidence. Every one of them is listed, with its source, on the numbers page.
Your answer is sealed before the reveal. Nobody can change it afterwards, including us. Thirty seals and you get your hit rate: do you know when you're right?
What we mark
The things a good consultation has that an exam can't see
Enough
The replay marks the moment you had enough to decide. Everything after it was for comfort, and it says so.
Before you look
Before any examination or test: what would each finding change? If the answer is nothing, we tell you before the patient does.
Holding risk
Over-investigating and over-referring are measured against the best call available to you, not against a list.
The net
Names, times, a way back, and whether the patient agreed. A vague net scores as vague.
Your hit rate
Every likelihood you commit is sealed before the reveal. After thirty, you find out whether you know when you're right.
Something's off
One tap when a story doesn't fit. We check whether it came before the alternative that mattered, and whether your plan changed.
The twist
Same case, days later, one thing changed. Did your call change for the right reason?
The return
Day three. They're back with the result. What you do with it is marked, because that is the job.
Also sound
Where two good GPs would differ, the feedback says so, names what it turns on, and never criticises what the evidence can't support.
How we mark
Every mark shows its working.
Not an AI's impression of your transcript. An exact solver works through every route the consultation could have taken, a second checker confirms it, and the replay shows your route beside the best one from where you were standing. Disagree with a mark and it gets recomputed, in the open.
Your route
- Asked the FeverPAIN questions 2 min · both present
- Back-up prescription
Best route from the same start
- Looked at the throat 2 min · severe, pus
- Asked the FeverPAIN questions 2 min · both present
- Antibiotic now
One thing to take away: look before you decide. The two-minute examination is what moves this patient across the line; the history alone leaves you at a back-up prescription.
The four verdicts: sound, also sound, depends (and on what), and can't say, where the evidence is too thin to criticise anyone and we say so instead.
379
tests pass on the marking engine, every commit
2,631
routes checked, twice, for the showcase case
0
marks given by a language model
1
thing to take away from each case. Not twelve.
Who it's for
Three people. One instrument.
If you're a trainee
The documented reasons people fail the SCA are unfocused questioning, vague plans and thin safety-netting. We mark all three, show you the fix, and give you one thing to change before tomorrow.
If you're a GP
Appraisal without the paragraphs. Three taps after clinic, twenty seconds of voice if you feel like it, and a reflective entry written from what you actually did. You approve it. You never write it. Coming in the beta.
If you supervise
One page that tells you what to talk about in the next hour, with the evidence behind every line and the replay one tap away. Free with any trainee's account.
Why not just another AI patient
It comes down to who does the marking
| Aspect | AI-patient simulators | Threshold |
|---|---|---|
| Who marks you | A language model's impression of the transcript, or a checklist | An exact solver, checked by a second one. No opinions. |
| What gets marked | Whether you covered the list | Whether each question was worth its minute, when you had enough, whether the call crossed the line, how good the net was |
| When you disagree | Nothing to push against | Dispute any mark. It is recomputed and the outcome is published. |
| When two good GPs would differ | One of them is marked wrong | Also sound. Or depends, and on what. Or can't say. |
| The patient | Can contradict itself | Can only say what is true in the case |
| The numbers | Undisclosed | Every one listed, with its source and its status |
Private beta · autumn 2026
Twenty trainees, their supervisors, one reviewed case family.
We want ST3s preparing for the SCA and the trainers who teach them. You get the daily ten minutes, the one page, and a say in what gets built. We get the only thing that matters: whether the replay changes what you do next.