Describe criteria for consistency: should identical inputs produce identical outputs?

The four moves for deciding when sameness is the point

A yes-or-no consistency question wants a pick, not a rulebook for every line Cloudline writes, so PICK does the work here.

One more thing before the walkthrough moves on: this pick covers one piece of copy, not the whole product. Most candidates hear "make it consistent" and answer by locking everything down. Say which piece earns the deterministic pass and which ones stay free to vary, and you've shown real judgment instead of reciting a safety word.

Let's learn

Cloudline is a small box behind Southridge Broadcasting's weather segment: feed it the raw storm data, and it writes the plain-English line that runs on the TV crawl, the radio script, the app push, and the website banner.

Before Cloudline, a producer typed that line by hand for each of the four channels, checking each one against the last to make sure the wording matched, about nine minutes a warning, longer once a real storm was already moving through.

With Cloudline, a producer gets all four drafts back in under twenty seconds. Every piece of copy it wrote, warning card included, ran through the model at the same setting, one built to keep the daily forecast from sounding like a recording: a sampling temperature of 0.7.

Here is the turn. That setting made the daily blurb sound fresh, and it worked fine there for years. But it also meant every warning card got drafted four separate times, once per channel, and each of those four calls was free to phrase the same fact a little differently. Nobody had ever asked whether the crawl and the app owed each other the same words.

At its worst, that looks like this: a tri-county severe thunderstorm warning goes out, and the TV crawl rounds the expiration time to 4:00 while the app states it exactly, 4:15, both drawn from the same National Weather Service alert. A viewer near the county line, watching the crawl, thinks the storm has already cleared by 4:00 and steps outside fifteen minutes early.

What I would leave alone. The daily forecast blurb, and anything else nobody reads two copies of side by side: internal drafts, archived past forecasts, and the seven-day outlook graphic, which nobody screenshots against a second channel to check if it matches.

The lesson. We built Cloudline to answer "does this sound like good forecast copy." It does. We never asked whether the same input, asked twice, owed anyone the same answer, and only the second question was the one that mattered on the one Thursday it counted.

The county line at 4:02

You don't need this to answer the question. Read it if you want to feel why the locked pass has to sit on the warning card before the next storm, not after.

Every storm season, Rutendo Chivayo checks the same three tabs from her desk at Southridge Broadcasting: the crawl, the app, and the radio script, one for each channel Cloudline feeds.

She's run product for Cloudline for five years, since before it wrote a single word on its own. Before that, a producer typed every warning by hand, once per channel, comparing each draft against the last so nothing shipped out of step. It took about nine minutes a warning, and there wasn't always nine minutes to spare.

For the first year Cloudline ran, Rutendo read all three channels side by side before anything went out, the way the old process used to force a producer to do. They always matched close enough. Routine watches, ordinary rain, nothing ever looked wrong.

So she stopped reading all three. She started skimming just the crawl, since that was the one most viewers actually saw.

Then she stopped skimming that too. If the pass check on the dashboard said CLEAR, she moved on to the next thing on her list.

Then, on a Thursday afternoon in April, a line of severe thunderstorms rolled toward three counties, one of them the county Southridge broadcasts from. Cloudline drafted the warning card the way it always did, once for the crawl, once for the app, each a separate call to the same model.

Same tool, two very different ways to disagree with itself

A new associate producer, two weeks into the job, messaged Rutendo directly: "quick one, the app says the warning ends at 4:15 and the crawl says 4:00, are those two different warnings?"

They weren't. Same county line, same National Weather Service alert, same input. One call had rounded the expiration time to the nearest quarter hour. The other had stated it exactly. Nothing was technically wrong. Nothing had ever been graded against the question of whether the two needed to say the same time.

Rutendo pulled up both channels at 4:02. The crawl still read "ends 4:00," already past its own stated cutoff, while the storm was still fifteen minutes from clearing the county line.

In the twenty minutes it took the assignment desk to sort out which time was right, the newsroom logged 46 calls asking the same question the new producer had just asked. Southridge read an on-air correction eight minutes later, stating the county's exact expiration time, word for word off the National Weather Service alert.

I want to say the problem is that one of the two channels was wrong. It wasn't. Both times were defensible readings of the same warning. The problem is that nothing had ever decided the two had to say the same thing, because nothing had ever been asked to.

Months earlier, when Cloudline first shipped, the team spent maybe ten minutes on the question of whether the warning card needed its own settings. Locking it down felt like extra engineering for a piece that had never once caused a problem. One shared setting, tuned for the blurb everyone thought about first, felt clean.

Rutendo pulled the warning card onto its own call the week after, locked to zero randomness, generated once and pushed identically to all four channels. Run the same April warning again, and the crawl and the app both say 4:15, because they're not two calls any more. They're one call, printed twice.

One design hands a viewer a fact she has to guess between. The other hands her the one fact that actually matters, said the same way twice.

And the thing I'd tell myself, if I could go back to the week we shipped Cloudline: we asked whether the writing was good. We never asked whether the same input, asked twice, owed anyone the same answer. On a calm Tuesday, it never mattered. On the one Thursday it did, we found out the hard way.

PICK, walked through on one warning card

This is a tradeoff about which piece of copy earns the deterministic pass, not a rule for every sentence Cloudline writes, so PICK is the tool.

Run PICK again, on a pharmacy counter

Anchorline Pharmacy runs the same question on refill alerts. Its AI drafts two kinds of message from the same fill record: a friendly "ready for pickup" text, and a "your refill window closes" text for controlled substances, sent by SMS and through the pharmacy's own app.

P. Lock the exact date and time in a controlled-substance refill deadline to one deterministic pass, reused identically across SMS and app. Leave the friendly "ready for pickup" reminder free to vary.
I. A patient who gets two slightly different "ready for pickup" texts loses nothing, maybe a shrug. A patient who reads two different refill-window dates for the same prescription might miss the real one, and for a controlled substance that can mean a lapsed prescription and a call back to the doctor.
C. The friendly reminder's mismatch is loud and cheap, caught by anyone glancing at two texts side by side. A wrong deadline is quiet until the day someone shows up to refill and finds out they were already a day late.
K. Drop the locked pass on any drug class whose refill window is wide enough that a same-day mismatch couldn't cause a missed dose, the same line Cloudline's team would use for a rule that no longer bites.

Flashcards (click a card to flip it)

Check yourself Score: 0 / 0