Design a confidence display for a medical information feature.

Verity Health runs Ask Verity, a symptom-information feature a patient can open any time of day or night. Emeka Sorensen has worked warehouse loading docks for nine years, and has used Ask Verity a handful of times before, for his kids' fevers and his own strains.

How to answer this, stage by stage

Nobody is grading whether your screen looks calming. They're grading whether it survives the one night the model's own confidence is wrong.

Let's learn

Ask Verity is a feature a national health information service runs. A patient describes their symptoms and gets a read on how urgent their situation might be.

Before Ask Verity, a patient called a nurse triage line, waited ten or fifteen minutes on hold during a busy night, and got a real person's careful judgment, thorough, but slow, and lines backed up badly during flu season.

Ask Verity replaced the second and third boxes with something instant. It hadn't yet replaced the honesty of the fourth.

Now Ask Verity gives an instant urgency read, day or night, no wait at all.

Here's the turn: for most cases, the label alone was enough, so people stopped reading past the color, because they never really needed to. Then a case comes along where the label's aggregate confidence underplays one dangerous, specific symptom combination, and by the time anyone reads the actual reasoning behind it, if reasoning was ever shown at all, the moment to act sooner has already passed.

At its worst, someone with a genuinely serious symptom combination stays home overnight because a single word, "Medium," was allowed to carry the entire decision.

Same night, same symptoms. Only what the screen was willing to say out loud changed.

What I would leave alone: the underlying symptom-matching model doesn't need a rebuild for the common, low-stakes cases, a cold, a mild sprain. Nothing here questions its accuracy on those.

The lesson: a confidence label is a summary, not an argument. The moment a summary is asked to carry a decision on its own, someone will trust it past where it was ever meant to go.

Now here is the same thing as a story

The short version above is what you'd say defending this design to Verity Health's clinical safety board. Read this one for how quietly the trust built up.

Ask Verity lives on a phone, and at 11pm on a Tuesday, Emeka Sorensen's phone was the only thing between him and a decision about his own chest.

Emeka has worked loading docks for nine years and knows the difference between a pulled muscle and something worse, most of the time. He'd used Ask Verity a handful of times before, for a kid's fever, a sprained wrist, and it always gave a fast, clear answer that matched what a nurse would have said.

Over those months, Emeka stopped reading past the colored band at the top of the screen, since it had never once been wrong for him before. There wasn't a specific bad case that changed things, more a slow trust that built up quietly, three visits, then five, each one confirming the color was all he needed.

After an unusually long shift, Emeka felt tightness across his chest and some shortness of breath climbing the stairs to his apartment. He described it to Ask Verity: tightness, mild breathlessness, and mentioned, almost as an aside in his own words, that his father had had a heart attack at fifty-two. The app returned "Medium urgency, monitor for 24 hours," the same band he'd seen plenty of times before for a strained back.

Four hours between a label he trusted and the moment he finally called someone.

Emeka went to bed, planning to call his doctor in the morning if it hadn't passed. It hadn't passed by 3am, and worried, he finally called a number a friend had once mentioned, an after-hours line. It turned out to be strain and stress, and a close look by an actual doctor confirmed his heart was fine. But it took a friend's memory of a phone number, not Ask Verity's design, to get him checked sooner than his own morning plan would have.

The old screen only ever showed the first of these four. Emeka never got to see the other three.

With the redesigned display, Ask Verity now shows "High urgency, seek care now" the moment the reported combination of chest tightness, shortness of breath, and early family heart history appears together, regardless of what the aggregate score alone would have said, and it shows exactly why: "your description includes chest tightness, breathlessness, and a family history of early heart disease. Together, these should be checked tonight, not in the morning." Run the same night forward: Emeka calls his doctor's after-hours line by 11:15pm instead of waiting until 3am.

The old design asked Emeka to trust a color. The new one shows him the actual sentence a nurse would have said out loud.

I built the label to be fast and simple, on purpose. It took watching how easily "Medium" can undersell a real risk to see that fast and simple isn't the same as honest.

SPARK, in one screenNot a lecture on making a screen feel reassuring. SPARK is what tells you which decision the whole design actually hangs on.

Only the top-right corner ever needs a hard floor rule. Everywhere else, the ordinary aggregate score is fine.

The recap, one line per letter: situation is a nurse triage line and a wait, payoff is teaching patients to read the reasoning instead of just the color, anchor is the redesigned display with its hard floor, risk is the night the aggregate score alone would have underrated a real combination, and keep out is holding back a specific diagnosis or a percentage on day one.

Each of these three is a real feature someone will ask for eventually. None of them belongs in the first version.

And if you want to be sure it really works, try it somewhere elseSame five letters, a home-insurance claim assistant instead of a symptom checker. A different anchor, a different hard floor.

ClaimClear is a home-insurance assistant that gives a homeowner filing a storm-damage claim a read on how likely it is to be approved. Lucia Bertrand filed a claim through it after a windstorm cracked her roof. Mapped onto SPARK: situation is a homeowner today, calling an adjuster and waiting days for a callback with no sense of where the claim stands; payoff is the habit to build, describing damage specifics fully instead of stopping the moment a likelihood number looks good.

The anchor here is structurally the same idea, aimed at a different failure: show which specific claim details are driving the approval-likelihood estimate, and hard-code certain causes, like pre-existing roof wear the policy explicitly excludes, to always require additional documentation, regardless of how high the overall approval-likelihood score reads. The risk ClaimClear's team designed against was a homeowner seeing "78% likely approved" and skipping the extra inspection photos that would have caught an exclusion clause early, only to have the claim denied weeks later once an adjuster finally looked closely.

Swap "symptoms" for "claim details" and "red flag floor" for "excluded cause floor." The same four parts still do the same job.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "show the reasoning behind the band, and hard-code known dangerous combinations to override the score," and stop.
Cost: there's no clinical panel available yet to define every red-flag combination. Say so honestly, and start with the small number of combinations already well established in emergency medicine, like this one, before expanding the list.
The model gets better, for real: if Ask Verity's underlying accuracy genuinely improves and false "Low urgency" reads become rare, that's still not a reason to remove the hard floor, a rare miss on a dangerous combination is exactly the one moment the floor exists to catch.

Where people run it wrong.
They treat a clean-looking single label as more trustworthy than a fuller explanation, when it's really just less checkable.
They let the aggregate score override a known dangerous combination, instead of the other way around.
They wait for a bad outcome to notice a design gap, instead of asking upfront what happens the one time the model is wrong.

How to use it live. When someone asks you to design a confidence display for anything that matters, ask yourself one question first: what happens the one night the model's aggregate score is confidently wrong? Design the anchor to survive that night, not just the ordinary ones.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0