Explain when a binary confident-or-not display beats a graded one.

Fenwick General Hospital's pharmacy uses Interax, an AI tool that scores possible drug interactions on a graded 0-to-100 risk scale. Priya Ilangovan is a staff pharmacist there. The same score, at launch, appeared both on her dashboard and on the after-visit summary handed straight to patients.

How to answer this, stage by stage

Nobody's grading whether you know graded scores are more precise. They're grading whether you know precision only helps the reader who can actually use it.

Here is what happens when the same number means two different things

Before Interax, before Priya's shift began, was a case where anything, from off-brand pain relievers to a supplement a patient forgot to mention, could combine badly. Fenwick's pharmacists cross-checked interactions by hand against a printed reference, about 6 minutes a prescription, across roughly 90 prescriptions a shift.

Same score, two very different amounts of power to act on it.

Interax scores a possible interaction in under a second, on a 0-to-100 scale, and displayed that same number both on Priya's dashboard and, at launch, on the after-visit summary every patient took home. The design team called it consistency. It was actually two different audiences reading one number meant for only one of them.

Here's the turn: the graded score wasn't wrong. It was accurate, well-calibrated, and genuinely useful, to the one reader trained to weigh it. The problem was handing that same precision to a reader with no way to check it, no training to interpret it, and no lever to ask a follow-up question before deciding.

The after-visit summary sits in the one corner a graded score should never be handed to: novice reader, high stakes.

At its worst, a patient reads a real, accurate risk score, misjudges what it means for their own specific situation, and doesn't call anyone, because the interface gave them a number to interpret instead of a decision to make.

What I would leave alone: Priya's own dashboard doesn't need this same simplification. She's trained to read a graded score, has the standing to escalate, and loses real information if the score gets flattened to a binary flag for her too.

The lesson: a number isn't neutral just because it's accurate. Who reads it, and what they can do next, decides whether that same number helps or quietly misleads.

Now here is the same thing as a story

The short version above is what you'd say defending this design to Fenwick's chief pharmacy officer. Read this one for how close the real incident came.

Before Interax, the best part of Priya's shift was never having to explain a printout to a worried patient at the counter. After Interax, that became the thing she explained most, patients holding a summary with a number on it, asking her what it actually meant.

The first months were fine. Pharmacists loved the graded score, using it to triage a shift's worth of prescriptions by how much attention each one needed. Nobody had thought much yet about the same number sitting on the sheet a patient carried out the door.

A woman in her sixties, managing declining kidney function, was prescribed a new blood thinner. Interax scored the interaction with her existing medication at 41 percent, printed plainly on her after-visit summary. She read it, reasoned that less than half wasn't so bad, and went home without calling.

The gap in the third box is the whole story. Nothing lived there for the patient to use.

She was hospitalized four days later with a bleeding complication, caught in time, but only because a family member noticed the symptoms and brought her in. When Priya's team reviewed the case, the score itself hadn't failed. The design handing that score to someone with no training to weigh it against her own condition had.

Four small facts replace one number nobody outside the pharmacy could actually weigh.

Fenwick didn't remove the graded score. They split the display: Priya's dashboard kept the full 0-to-100 scale with contributing factors listed. The patient summary now shows one of two things: "No action needed" or "Call your pharmacist before your next dose," with a single plain-language reason underneath either one.

GUARD, said plainly, no moralizingNot a lecture on health literacy. GUARD is what tells you exactly which reader the precision is actually for.

Either sign, alone, is worth switching a screen to a binary flag instead of arguing over exact thresholds.

The question is never "which display is better." It's always "who's reading it, and what can they actually do next."

The recap, one line per letter: groups is naming Priya and the patient as two separate readers with two different amounts of power, unequal is the harm concentrating on patients whose exceptions a general score can't see, ability to contest is the whole reason binary wins for patients, reduce is the actual split-display fix, and detect is watching real callback behavior, not just whether the redesign looks cleaner.

And if you want to be sure it really works, try it somewhere elseSame five letters, a county probation office instead of a hospital pharmacy. A different kind of unequal breaks the second story.

Calder County Probation uses Reliant, an AI tool that scores a defendant's recidivism risk on a graded percentage, shown to both the supervising officer and the defendant during case review. Deacon Marsh, a probation officer there, ran the same GUARD analysis on Reliant's shared display. Mapped onto GUARD: groups is the officer, who has real discretion to adjust supervision conditions, and the defendant, who receives a supervision plan built partly from the score; unequal is that a defendant with a borderline score, say 58 percent, gets the least clear treatment of anyone, since the number reads as almost arbitrary and gives them nothing concrete to respond to.

The ability-to-contest step is where this story diverges hardest from Priya's. A defendant can formally appeal a supervision decision. They cannot appeal the number itself, since there's no defined process for challenging how a percentage was calculated, only for challenging what was done because of it. The reduce step: Reliant now shows officers the full graded score with contributing factors, and shows defendants a binary "standard supervision" or "enhanced supervision, here's why" with the specific named factors, not a percentage.

Swap "call your pharmacist" for "here's who to talk to about your supervision plan." The shape of the fix travels.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "graded for the trained reader who can act on it, binary for the reader who can only guess," and stop.
Cost: there's no engineering time this sprint to build a second display. Say so honestly, and hide the graded score from patient-facing views entirely first, since removing a misleading precision costs less than building a new one.
The model gets better, for real: if Interax's scoring genuinely gets more accurate, that's still not a reason to show patients the raw number. It's a reason to trust the binary flag's threshold even more, since a better model makes the flag itself more reliable.

Where people run it wrong.
They treat "showing the same number everywhere" as fairness, when it actually hands precision to a reader who can't use it.
They add a binary flag but leave the graded score visible right next to it, so patients read the number anyway.
They build the flag but skip the escalation path behind it, leaving someone with nowhere real to go when it fires.

How to use it live. When someone asks whether binary or graded is better, ask back: can the person reading this number push back on it, ask a follow-up, or act on the nuance? If not, binary wins, no matter how accurate the graded score is.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0