Argue that trust is the only quality metric that matters, then argue against it.

Say trust is the only quality metric that matters, and you are half right. The other half is what trust is standing on when nobody is watching.

How to answer this, stage by stage

Nobody's grading whether you can argue passionately for trust. They're grading whether you can argue against your own case just as hard, then still land on one answer.

Let's learn

Picture two arguments about the same product, both true on their own, and only one of them safe to build a roadmap on.

Larkline is a phone service. Once a day, it calls an elderly person living alone, asks a few short questions, and sorts the call into one of three outcomes: fine, flag for a family callback, or call emergency services right now.

Before Larkline, a home care agency did these calls by hand. One caller could get through about 60 people in an eight hour shift. On a day with more than 60 people on the list, and there almost always were, the last few calls got skipped, and nobody found out until the next morning's check.

Now Larkline calls everyone on the list, every day, in under 90 seconds a call. Checked every week against 1,000 nurse-reviewed calls held back just for grading it, Larkline gets the right outcome about 91 percent of the time.

Nine percent wrong sounds like the whole story. It isn't. The real damage came from one wrong answer, and what a family did after it.

One caller, an 84 year old man living alone, told Larkline he felt "a little dizzy today, nothing new," three mornings running. That's a symptom shared by dozens of harmless causes, and by an early stroke. On the third morning, Larkline logged it as fine, call again tomorrow. Nobody flagged his daughter. That evening he collapsed. A neighbor found him fourteen hours later. He survived. His recovery lost ground those fourteen hours can't give back.

At its worst, a call service families stop trusting is worse than no service at all. The old agency, slow as it was, called everyone it had time for and never claimed more than that. Larkline, once families stop believing its "fine," gets skipped exactly on the mornings it matters most, because a family that's already calling their parent directly has stopped reading the app's verdict at all.

What I'd leave alone: the routine tier, calls like "did you take your morning pills," genuinely doesn't need this. Warm, confident phrasing there costs nothing when it's wrong, someone just gets called again tomorrow. Spending audit time matching confidence to accuracy on routine calls would take time away from the tier where a miss actually costs something.

The lesson: a number can be completely honest and still be the wrong number to chase. Ninety one percent accurate told the team the model was fine. It never told them whether the confidence in its voice matched the accuracy underneath it, on the one tier where that gap has a person's name on it.

Now here is the same thing as a story

Read the long version below when you want to feel the difference between the two failures, not just be told which one wins.

The paper log Tarquin Naismith keeps beside his monitor has a torn corner, worn soft from three years of thumbing back through it during a call he doesn't like the sound of.

Before Larkline, Tarquin ran the phone room at a home care agency for six years. He could hear a real problem in someone's voice inside the first ten seconds, tone, pace, the pause before "I'm fine." A model doesn't have ten years of ears. He knows exactly what that costs.

Larkline launched with Tarquin running quality on it. For the first year, Monday mornings were easy. He'd pull the week's calibration report, a nurse's read on 200 calls compared to Larkline's own, and it came back somewhere between 89 and 92 percent, same as the month before. He'd read it, note it, and get back to his coffee before it went cold.

When the flag and emergency tiers first got their own weekly slice of that report, eighteen months back, Tarquin pulled it open every single Monday for two months. Volume on the emergency tier was tiny back then, maybe a dozen calls a week, and the tier number never once said anything the blended one hadn't already said. He started skipping it some weeks, when Mondays got busy. By month nine, he'd stopped opening it at all.

It came back on a Thursday, not a Monday. A message from a neighbor, forwarded twice before it reached him: an 84 year old man, found on his kitchen floor, fourteen hours after Larkline told him he was fine.

Half the room wanted to retrain the model that afternoon. Tarquin asked for the afternoon instead, to pull every emergency and flag tier call from that week, all 31 of them, not just the slice the weekly sample had happened to touch.

It was never really about whether the blended number moved. It hadn't. What moved was whether a weekly sample, drawn the same way regardless of what each tier costs to get wrong, could actually tell Tarquin the emergency tier was safe. It couldn't. It never had.

The decision that opened the door went back to the week the tiers were split out. Someone asked, in passing, whether the emergency tier needed its own full review instead of riding along in the blended sample. The answer was no, a dozen calls a week didn't justify a second report. Nobody came back to that question as the tier's stakes grew, only its count.

Run the same Thursday again with one change: every flag and emergency tier call gets reviewed in full each week, not sampled, about 25 calls, twenty minutes of a nurse's time. The same words, "a little dizzy, nothing new," still get logged as fine by the model on day one. But the pattern, three mornings running, the same phrase, gets caught in that Monday's full review, two days before it would have mattered, not fourteen hours after.

One design trusted one sample to answer two very different questions. The other asks each tier the question its stakes actually deserve.

What I'd tell myself, back in that meeting when tiers were first split out: the moment a metric starts covering cases with wildly different costs, ask whether one sample can still answer for all of them, or whether it only ever could because the expensive case was too rare to notice yet. Nobody asked. That's on the room, not on the model.

Five ranks, since both sides had a point

This isn't a coin flip between two good essays. It's ORDER, run on the same tension, until one side genuinely outranks the other.

Both cost you something. Only one of them ever pays it back.

Skip the first box and the other two are just hope.

Three things worth stating directly, since this is where the real judgment sits. The alternative the team rejected right after the missed call was making Larkline's whole voice more hedgy and cautious on every call, never sounding too sure about anything. It lost because the emergency and flag tiers are a sliver of total volume, and softening every routine call to patch one tier's floor would have made the other 95 percent of calls worse for families who never had a floor problem to begin with. The AI specific failure mode worth naming by name is confidence outrunning accuracy on an ambiguous symptom cluster, a phrase like "a little dizzy, nothing new" that's genuinely shared by a harmless cause and a dangerous one, where the model's voice stayed just as certain regardless of which one it actually was. The guardrail is two part: a weekly calibration audit checking whether expressed confidence matches measured accuracy, tier by tier, and a named phrase list that always escalates to a human review, whatever the model's own confidence says. That guardrail isn't free. Tightening the escalation threshold on ambiguous clusters sends roughly 60 more calls a week to a human callback instead of closing them as fine, at about 12 minutes of a nurse's time each, a real cost accepted on purpose, only on the tiers where a miss is this expensive. And the bar that decides whether a tier is safe enough was never zero misses, a system running thousands of calls a day can't promise that on a probabilistic call. It's an audited floor, checked in full every week on the tiers small enough to check in full, not estimated from a random slice too small to trust.

And if you want to be sure it really works, try it somewhere else

Same five ranks, a warehouse floor instead of a phone line, nothing about elder care anywhere in sight.

SiteGuard is a camera system that watches a warehouse floor and flags near misses between forklifts and people on foot. Reuben Ondrejka is the safety manager who decides how those flags get handled.

The case for trust alone: workers stop trusting an alert system fast. SiteGuard's false alarm rate spiked to 40 percent for one week, a glare problem on two cameras, and floor staff started waving off every alert that week, real or not. Even after the glare got fixed and false alarms dropped back to 4 percent within days, staff kept ignoring a chunk of real alerts for weeks after.

The case against it: an earlier, more cautious build of SiteGuard almost never alerted at all, and staff loved it, nothing ever cried wolf. It felt trustworthy. It was also missing most of the real near misses, quietly, because a system tuned to rarely speak is a system tuned to be wrong in silence. Nobody found out until a real near miss went unflagged and a worker was one second from being hit.

Quiet and correct look the same from the floor. Only a recall audit tells them apart.

Same rank as before: trust first, but only once the alert has a provable recall floor on real near misses underneath it. A quiet system and an accurate system aren't the same thing, and only the audit tells you which one you've built.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the rank, trust first, only on a proven floor, and name the one audit that checks it.
Cost: there's no budget this quarter for both the calibration audit and a warmer, more reassuring voice update. The audit wins every time, warmth on a broken floor is just decoration.
The model got better, for real: say overall accuracy improved this quarter. That's not proof the worst tier improved with it. A model can get better on average while the tier that costs the most stays exactly as blind as before.

Where people run it wrong.
They read a rising average as proof there's no gap anywhere, and never check the worst tier on its own.
They fix a trust dip by making the voice warmer and more certain, instead of fixing the floor underneath it.
They patch a miss quietly, thinking silence protects trust, when a disclosed, bounded failure is what actually earns it back.

How to use it live. Say the real tension out loud before answering it: "is this asking me to pick a metric, or to rank two different ways trust actually breaks." That buys a beat to think instead of guessing out loud in front of the interviewer.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0