What is the design response to a model that is well-calibrated versus poorly calibrated?

Corrigan Bottling runs six filling lines at a contract beverage plant. Adaeze Umeh is the plant's Reliability Engineer. Ferroscope is the AI vendor tool that watches motor and bearing sensors and gives a number: how likely a part is to fail in the next 48 hours.

How to answer this, stage by stage

The interviewer isn't checking whether you know what "calibration" means. They're checking whether you'd treat one confidence number the same on every line, or notice that the same digit can mean two different things.

Let's learn

What happens when a machine tells you the truth on day one and starts lying to you, politely, eight months later?

Ferroscope reads vibration and heat sensors on bottling-line motors and tells the team how likely a bearing is to fail soon. Before it existed, Corrigan Bottling serviced every motor on a flat 90-day schedule, whether it needed it or not, and still ate 14 unplanned stoppages a year, about six hours each. On the two original lines, once Ferroscope launched, unplanned stoppages fell to about 2 a year, each one flagged days ahead.

The number never changed shape. What it was actually measuring did.

Here's the turn: eight months later, three new lines were added, running slightly different motors Ferroscope had never seen in its training data. The model kept producing the same kind of number, a clean percentage, on those lines too. Nobody had checked whether that number still meant what it used to mean.

At its worst, this doesn't just mean a few false alarms. A confident-looking, low number can also tell an operator a bearing is probably fine when it isn't, which is worse than any noisy alarm, because nothing on screen tells them to double-check it.

What I would leave alone: Lines 1 and 2 don't need this fix at all. Their number has two years of real outcomes behind it, and it's still right about as often as it claims. Hiding a number that's actually earned its trust would just slow the team down for no reason.

The lesson: a model's honesty isn't a fixed trait, it's a property of the specific slice of the world it's being asked about right now. A design has to be able to say "I haven't checked this one yet" instead of always sounding equally sure.

Now here is the same thing as a story

The short version above is what you'd say out loud in the interview. Read this one for how close the actual near miss came.

Adaeze Umeh can tell, from the sound of a filling line alone, which motor is about to give her trouble. Nine years on the floor will do that. When Ferroscope went in on Lines 1 and 2, she watched it catch two real bearing failures in its first quarter that the old 90-day schedule would have missed completely.

The good months were quiet ones. By 6 a.m. most Tuesdays, Adaeze's team had already cleared the overnight flag list, usually one or two low numbers worth a glance, nothing urgent. The floor trusted the screen the way they trusted a coworker who was almost always right.

Eight months in, three new filling lines came online, using a slightly different motor mount Ferroscope had never trained against. The flags kept appearing, worded exactly the same way. Nobody on the floor had a reason to treat them differently. Why would they?

Four flags, same stated confidence. Only two of them had actually earned it.

On a Thursday night, a bearing on Line 5 threw off a reading. Ferroscope's flag came back low, a reassuring number that read like "probably fine, check it next week." The old 90-day schedule would have had a technician's hands on that exact bearing within four days regardless. Nobody escalated. The screen looked calm.

Adaeze's plant ops director pulled ten flagged predictions at random the following week, from three different lines, and checked each one against what had actually happened. On the two original lines, the numbers held up close to what they claimed. On the two newest lines, of the flags stating 80% confidence or higher, only about 38% had actually come true. The badge had been quietly wrong for months, and it had never once said so.

Adaeze's team could go check the model's math after the fact. The operator on shift at 2 a.m. never could.

Adaeze's redesign split the display in two. Verified lines kept their plain number. Unverified or drifting lines lost the number and got the raw evidence instead: which sensor tripped, by how much, and a real hit rate built from that specific line's own history, however thin it still was.

Same digits on screen. One had two years of proof behind it. The other had none.

GUARD, the whole page in one lookNot a fairness lecture. GUARD is what tells you which number is allowed to sound confident.

One question decides the whole display. Everything else follows from the answer.

One box in the middle was empty the whole time. Nothing there ever asked whether the number could be trusted yet.

The recap, one line per letter: groups is the team that can check versus the operator who can't, unequal is the newest lines carrying the worst risk, ability to contest is the operator with no way to verify the number themselves, reduce is splitting the display by calibration status, and detect is a rolling per-line accuracy check with an automatic alert.

And if you want to be sure it really works, try it somewhere elseSame five letters, a crop-disease detector on a farm instead of bearing sensors on a bottling line. A different subject carries the risk this time.

Ridgeback Growers Cooperative uses an AI tool that scans leaf photos and estimates how likely a crop is infected with a fungal blight, so an agronomist named Zuriel Wanjiru can decide which fields need spraying this week, not next. The tool launched trained on the cooperative's three original crop varieties and was solidly calibrated on all three. Then the co-op added two new varieties this season, grown from different seed stock the model had never seen.

Mapped onto GUARD: groups is the co-op's agronomy team, who can walk a field and double-check by eye, versus the smaller growers who lease equipment and trust the app's number outright because they can't afford the agronomist's time. Unequal is that the new varieties, again, carry the worst miscalibration: stated high-risk flags on the three original varieties were confirmed 82% of the time, but on the two new ones that fell to about 51%, close to a coin flip. Ability to contest is the leaseholder growers, who have no field history of their own to compare the number against. Reduce is the same fix: show the plain infection-probability number only for the three original varieties, and for the two new ones, show the raw leaf symptoms the model actually detected, plus how many total scans of that variety exist so far.

Same shape of gap. A different subject carries it this time: the grower with no agronomist on staff.

The old decision here isn't quite the same one. Ridgeback's team hadn't picked one universal display style out of convenience, they'd made a packaging decision: the app charged growers per scan, which pushed leaseholders to save their scans for the crops they were most worried about, meaning the new varieties, exactly the ones with the least calibration data, got scanned the most by the people least able to double-check the result.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "split the display by whether the number has been checked; show it plain where it's earned, show the evidence instead where it isn't," and stop there.
Cost: there's no budget to rebuild the display right now. Say so, and start by just logging predicted-versus-actual per line for free; that alone tells you where the real risk sits before you spend anything on the fix.
The model gets better, for real: if Ferroscope is retrained and a new line's calibration checks out clean for a full quarter, that line graduates back to showing the plain number, the same rule running in reverse.

Where people run it wrong.
They treat "the model is usually accurate" as the same thing as "the model is calibrated everywhere."
They let a single global display setting quietly outlive the one line it was actually tested on.
They wait for an incident to reveal drift instead of tracking the predicted-versus-actual gap on their own.

How to use it live. When someone asks how you'd handle a well- or poorly-calibrated model, ask back: has this exact slice of the model's job actually been checked against real outcomes recently? Let the honest answer, not the model's confidence, decide what the screen is allowed to say.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0