Describe the tradeoff between showing sources and cluttering the interface.

Dunmore Falls' building department runs ZoneCite, a chat tool on its permit page that answers zoning questions and cites the specific code section behind each answer. Renata Sowinski leads permitting technology for the department.

How to answer this, stage by stage

The interviewer isn't asking you to pick a side and defend it forever. They're asking whether you can name which kind of wrong actually hurts someone.

The chat box in the corner of the permit page

The chat box sits at the bottom right of the city's permit page, small, easy to miss, and busier than anyone on staff expected. Before ZoneCite, residents called or waited in line to ask "do I need a permit for this," and staff answered from memory or looked the code up by hand, about 8 minutes a question, across roughly 60 questions a day.

Most answers should feel like the receipt. A few genuinely need to feel like the brief.

ZoneCite's first version answered instantly, but under every response it printed the full statute text and every cross-reference, treating every question, "what are your office hours" included, like a legal filing. Average time to read a full answer stretched to 14 minutes, longer than the old phone call.

Here's the turn: the fix for clutter isn't removing the source. It's changing how loud it is. Dunmore Falls' second attempt swung the other way, stripping citations entirely from any question that sounded routine, and that swing is what actually let a wrong answer through.

One of these costs a shrug. The other costs a torn-down deck. They are not the same kind of wrong.

At its worst, an under-sourced answer doesn't just annoy someone later. It lets a resident act, in good faith, on a clean-looking answer that quietly left out the one detail that mattered, with no citation anywhere to have caught it.

What I would leave alone: a question like "what are your office hours" or "where do I submit a paper form" genuinely doesn't need a zoning citation. Nothing about those answers depends on a code section, and adding one there would be clutter with no offsetting benefit.

The lesson: clutter and missing sources aren't opposite ends of one dial you can just split the difference on. They're two different failures, and the fix for one is not automatically safe for the other.

Now here is the same thing as a story

The short version above is what you'd say defending this design to the city council. Read this one for how the actual near miss unfolded.

Renata had spent four years modernizing Dunmore Falls' permit office before ZoneCite, replacing paper logs with a searchable system nobody else in the building had bothered to build. She knew the zoning code well enough to catch a wrong answer by ear.

ZoneCite's first months were good. Residents got instant answers, staff got fewer repetitive calls, and everyone assumed the dense footnotes underneath, while a little much, were at least thorough. Nobody complained about missing information. They complained about reading time.

So the team simplified. Any question ZoneCite classified as routine, a wide majority of them, got a clean answer with no citation block at all, on the theory that citations were for complicated cases, not easy ones. Nobody decided this was risky. It looked like a reasonable cleanup at the time.

The middle branch is where the near miss lived: a question that looked routine but sat inside an overlay district.

A homeowner asked whether he needed a permit to build a small backyard deck. ZoneCite classified the question as routine, since decks under a certain size usually don't need one, and answered cleanly: no permit required. It didn't mention that his specific lot sat inside a floodplain overlay district, where the same deck did need one.

He built the deck over a weekend. The overlay-district requirement surfaced three months later during an unrelated inspection next door, and the city required the deck brought up to code, at his own cost, well after it was finished.

Dunmore Falls didn't go back to the wall of footnotes. They built the collapsed source strip instead: one tappable line under every answer, expanding automatically on any question touching permit requirements at all, routine-sounding or not.

Four small facts, collapsed by default, expanded on a single tap. Neither a wall nor a blank.

PICK, said out loud before the reasonsNot a debate about clutter in general. PICK is what tells you exactly where the line sits.

Either failure is a real failure. Neither one is the safe default.

The recap, one line per letter: position is the collapsed strip that expands on stakes, impact is naming both the bored skimmer and the burned homeowner, cost asymmetry is why a missing citation costs more than an ignored one, and kill criteria is letting real expand rates move a question type between tiers.

Two overcorrections and a fix, five months apart, before the screen finally found the actual line.

And if you want to be sure it really works, try it somewhere elseSame four letters, a public library's reference desk instead of a permit office. A different impact breaks the second story.

The Dunmore Falls public library runs AskStax, an AI reference tool that answers patron research questions and cites the encyclopedia entries, databases, or archived articles behind each answer. Halvard Bruun leads reference services there. Mapped onto PICK: position is the same collapsed source strip, expanding automatically whenever a patron's question sounds like it's for a paper or a citation, not a quick personal curiosity; impact is different from Dunmore Falls' permit case, since the person hurt by a missing source here isn't risking a fine, they're a student who can't cite an unsourced claim in their own bibliography and has to redo the research from scratch.

The cost asymmetry still holds, just in a different currency. A patron skimming a quick fact loses nothing by not tapping a citation. A student building a paper on an AskStax answer with no traceable source loses the whole citation, and sometimes the credibility of the paper itself if a professor asks where the fact came from. The kill criteria are the same shape too: if "where's this from" follow-up questions rise for a topic area, that topic's answers move to auto-expanded sourcing.

Swap "code section" for "database entry," and the same four-part card does the same job.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a one-line source strip by default, full citation auto-expanded on high stakes," and stop.
Cost: there's no time this sprint to build tiered auto-expansion logic. Say so honestly, and ship the flat one-line strip on everything first, since even a always-visible-but-collapsed citation beats either extreme.
The model gets better, for real: if ZoneCite's classification of "routine" gets more accurate over time, that's still not a reason to remove the citation entirely from any tier. It's a reason to trust the collapsed default more, not to go back to hiding sources altogether.

Where people run it wrong.
They treat "add sources" and "reduce clutter" as opposite ends of the same dial, instead of two separate failures with two separate fixes.
They classify a question as low-stakes based on how it sounds, not on what's actually true about the specific case.
They ship a fix for one side of the tradeoff and declare the problem solved, without watching whether the other side quietly got worse.

How to use it live. When someone hands you a "sources versus clutter" question, ask back: for this exact answer, is the source's whole job to be skimmed past, or to be found when someone really needs it? Design the loudness of the citation around that answer, not around a general rule for the whole screen.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0