Design the transparency layer for a product I describe.

Here is the product: Harrow Telecom built TraceLine. A field technician stands at a broken piece of network gear, TraceLine reads the equipment's signal data, and it tells the technician what it thinks is wrong. Priya Vashisht has fixed roadside cabinets for six years and can usually guess a fault before her meter confirms it.

How to answer this, stage by stage

You're not being graded on whether you remember every SPARK letter. You're graded on whether you can point at one design decision and defend it.

Let's learn

Harrow Telecom built TraceLine, an AI tool that reads the signal data coming off a broken piece of network gear and tells a field technician what it thinks caused the fault.

Before TraceLine, a technician like Priya Vashisht climbed up to a roadside cabinet, tested the line with a handheld meter, and paged through a forty-page paper fault chart to match the pattern to a known cause. That took about 47 minutes, on a good day.

An early version of TraceLine cut that 47 minutes down to 18. It read the signal and printed one line: the most likely cause, and a percentage next to it.

This is what TraceLine had to replace: four slow, manual steps, each one a place a technician's own judgment used to sit.

The turn. The extra minutes TraceLine saved were never the real problem. The problem was what technicians stopped doing with those minutes: checking the signal themselves before opening a panel.

At its worst: a technician would read "Corroded connector, 91%," swap the connector, and drive off without checking the actual line. Two weeks later, a second truck roll to the same cabinet, because the real fault was a cracked seal the first score never mentioned.

What I would leave alone: the paper fault chart, kept as a backup for cabinets that emit no clean signal at all, usually older rural sites. No explanation card, however well designed, can explain data that was never there to begin with.

The anchor is the third box. Showing the cause TraceLine ranked second is what gives a technician somewhere real to look when the first one is wrong.

The lesson: a single number on a screen isn't a neutral fact. It's an invitation to stop checking, and a person who's good at their job will always take a fair invitation to work a little less.

Now here is the same thing as a story

The short version above is what you'd say in the room. Read this one for how the card actually got built.

Priya Vashisht has worked Harrow Telecom's northern routes for six years. Before TraceLine, she could often name a fault from the sound the cabinet fan made when she opened the panel, and she'd only pull out her meter to confirm it.

TraceLine's first months were good ones. Her average visit dropped from 47 minutes to 18. She started closing tickets before her lunch break instead of after it.

The habit thinned out in three beats. First, she stopped double checking cases the score called "high confidence." Then she stopped opening the paper chart at all, even for the odd ones. Then, without deciding to, she stopped touching her meter before swapping a part TraceLine named.

The trigger was small: a colleague on the southern crew, Jorik, mentioned over coffee that his second truck roll count had crept up all quarter. Nobody had connected it to TraceLine yet.

Same wrong top guess, two designs. One sends a technician up the ladder twice. One catches it on the ground.

Priya pulled her own numbers that week. Wrong-part swaps, cases where the part she'd replaced turned out not to be the actual fault, had climbed from about 6 percent of her visits to 21 percent over twelve weeks.

Nobody decided to stop checking on any single day. It happened between week one and week six, and nobody wrote it down.

She raised it at the next regional call, not as a complaint about TraceLine's accuracy, since the model itself hadn't gotten any worse, but as a question about what the screen was actually asking her to do.

The redesigned card shipped in week fourteen: signal pattern, top cause with a confidence band, and the cause TraceLine ranked second. The team considered a stricter fix first, a mandatory second confirmation step on every single recommendation, and rejected it. It would have erased the 29 minutes TraceLine saved on every visit, which defeats the reason the tool exists.

None of these four were wrong ideas. They were just riskier than the team needed to be on day one.

Same fault, same cabinet, replayed with the new card: Priya sees "Corroded connector, 68 to 79 percent" next to "Cracked seal, second most likely." The band is narrow enough to make her check the seal before she opens her toolbox. Four minutes added to her visit. Zero return trips.

I built the single-score version because it tested well in a demo and looked clean on a small screen. It took watching Priya's own wrong-dispatch count nearly quadruple, on a model that hadn't changed at all, to see that the card was never really about looking clean. It was about giving her a reason to keep doing the thing she was already good at.

SPARK, in one screenNot a story about a mistake. SPARK is what forces the design to survive the day it's wrong, before you ship it.

The two pieces of the card sit safely in the top left. Auto-apply and the full dashboard sit in the corner that earns the "not yet."

The recap, one line per letter: situation is Priya's 47-minute manual diagnosis, payoff is reading the reasoning instead of the label, anchor is the three-part card, risk is the wrong-dispatch spike the score-only version caused, and keep out is the auto-apply mode the team deliberately didn't ship.

And if you want to be sure it really works, try it somewhere elseSame five letters, a disaster-relief nonprofit instead of a telecom crew. A different flip family shows up here too: this one is about a delegation risk, not over-trust.

Bridge Relief Network runs IntakeLine, an AI tool that reads a family's shelter-intake answers and flags urgent needs for the coordinator on shift. Esteban Duarte manages intake at one of their busiest sites, and used to interview every arriving family himself before IntakeLine existed.

Mapped onto SPARK: situation is Esteban assessing each family by hand on paper forms, asking the same trauma-sensitive questions over and over across shifts. Payoff is coordinators trusting the shared record enough to stop re-asking those questions at every handoff. Anchor is that every auto-suggested flag, like "high priority: medical need," shows the exact answers that triggered it, with the underlying fields editable, not just an accept-or-reject button. Risk is the day IntakeLine misses something because a family answered indirectly; the anchor design lets a coordinator see the raw answers behind the flag and catch the gap immediately, instead of trusting a flag that quietly said nothing was wrong. Keep out is any automatic bed or resource assignment straight from the flag. A person still makes that call.

Three branches, and none of them let a flag become a decision on its own. A person always closes the loop.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "show the reasoning, not the label, so trust can be partial," and stop.
Cost: there's no budget this quarter for a full redesign. Say so honestly, and ship the rejected alternative first; it's the cheapest single addition and it catches most of the close calls on its own.
The model gets better, for real: if TraceLine's overall accuracy improves next quarter, that's not a reason to remove the card. A better model still has a worst 5 percent of cases, and those are exactly where a technician needs the reasoning most.

Where people run it wrong.
They treat "transparency" as a disclaimer added after the design is done, not a decision made before it.
They assume more information always helps, and bury the one useful signal, the rejected alternative, under ten fields nobody reads.
They design for the demo case, where the model is right, instead of the case that actually matters: the one where it's wrong.

How to use it live. When someone asks you to design a transparency layer, ask yourself one question first: what does this person do the day the model's top guess is wrong? Design backward from that answer, not forward from what looks clean on a screen.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0