Contrast the definition-of-done for a CRUD feature and for a retrieval feature.

Marigold reads a user's real bank transactions and answers real questions about their money, in plain words. This sprint it ships two features: Bill Tracker, where you manually add and edit recurring bills, and Ask Marigold, where you type a question and it reads your own transaction history to answer it. Talitha Havercombe owns what "done" means for both. Three weeks after launch, a user named Theron Kilgannon finds out the two checklists were never supposed to be the same one.

How to answer this, stage by stage

Nobody is grading whether you know the word "retrieval." They are grading whether you would let the same checklist wave through a save button and a real-money answer.

Let's learn

Marigold is an app that reads your real bank transactions and tells you, in plain words, whether a money move is safe.

Before Ask Marigold shipped, a user like Theron Kilgannon had one way to check a decision like this: open Bill Tracker, scroll his own categories, and do the math himself. It took him about twenty minutes, and he trusted his own math, because he had checked it himself.

One word, done, stamped on two features that fail in completely different ways.

Ask Marigold changed that. Type a real question, get a real answer, in about four seconds. In testing, it answered 47 of 50 questions the engineering team wrote for it correctly. Ninety-four percent. Everyone called it done, the same way Bill Tracker got called done: one checklist, one QA pass, ship to a small rollout.

Here's the turn. That extra six percent was never the real risk. The real risk was what happens the one time it is confidently wrong about a question with real money behind it, and nobody built a way to catch that before a user acted on it.

Nothing about this failed loudly. The search just never checked the one category Theron had made up himself.

Theron had $3,200 sitting in his rent buffer. He asked Ask Marigold whether he could move $2,000 to cover a car repair without hurting his rent goal. Its retrieval searched his standard spending categories and never surfaced a $2,150 annual insurance premium he had filed under a custom label, "House stuff," months earlier. Ask Marigold said yes: moving $2,000 would leave him a solid $1,200, "comfortably covers your typical monthly costs." He moved the money.

What it costs at its worst: a real decision, with real money, based on a wrong answer stated as a plain fact. There's no undo button on a car repair already paid for, or a rent payment that lands two days late. Theron paid a $35 overdraft fee and a $75 late fee from his landlord, on top of being caught $950 short. And he stopped trusting Ask Marigold with anything that mattered.

What I would leave alone: the notification toggle, the sign-in screen, the export-my-data button. None of them have a model guessing at anything. The old one-checklist bar is still exactly right for those, and building a heavier bar for them would waste the two weeks it actually takes to build one properly.

The lesson: a checklist that works for a save button does not automatically work for a feature that has to guess. If the same word, done, is asked to mean two different things, it needs two different ways of being earned.

Now here is the same thing as a story

The short version is above, for when you're in the room. This one is for feeling why a rising test score and a real, defensible answer are not automatically the same claim.

Talitha Havercombe has run launch reviews at Marigold for four years. She built the checklist everyone uses: a one-page form, spec matches build, QA signs off, no P0 bugs, ship to ten percent. It has caught almost every real problem that ever mattered, because almost every feature Marigold ever shipped was a form, a save, a toggle.

Bill Tracker went through her checklist the normal way. An engineer clicked through every field, tried a negative number, tried a decimal, tried deleting a bill mid-edit. It passed in an afternoon.

Ask Marigold went through the same form. Underneath the usual boxes, the team had added one line of their own: a fifty-question set they had written together over a week, covering the kinds of things they figured a user might ask. "How much did I spend on groceries." "What's my biggest bill this month." They ran it, and 47 came back right. Talitha signed off. It felt no different from signing off on Bill Tracker. A number, a bar, a pass.

For three weeks, on the ten percent of users who had it, nothing went wrong. The team watched their own dashboard: response time held, nobody filed a bug, the golden set still scored 94 percent every night. It looked exactly like a feature that was fine.

The same desk. The same checklist. A very different Tuesday.

Then support forwarded her a message. Not a bug report, a support agent had written across the top: "This one's not a bug, I think, but you should read it." Theron had written three paragraphs about asking Ask Marigold if he could afford a car repair, being told yes, and then watching his account go $985 negative four days later when an insurance bill he'd forgotten about landed on top of it.

Talitha pulled the transcript. Ask Marigold had told him, plainly and confidently, that $1,200 comfortably covered his typical monthly costs. She pulled his transaction history next. The $2,150 premium was right there, filed under a category he'd made up himself eight months earlier, "House stuff." The retrieval had searched his standard categories and never once looked inside his own.

The decision Talitha would take back happened in a planning meeting two months earlier, when the team decided how Ask Marigold would earn the word done. Someone suggested writing test questions themselves, since building a real question set from actual users would take weeks they didn't have. It felt efficient. It felt like moving fast on a feature everyone was excited about. Nobody in that meeting asked the one question that would have caught this: would a person who didn't build the retrieval ever think to ask it that way?

She ran the numbers again, properly this time. Forty real question patterns pulled from the three weeks of rollout logs, graded against a new rubric: did the retrieval find the right transactions, did the stated number match, did it hedge when the picture was incomplete. Twenty-four of forty cleared the bar. Sixty-one percent, not ninety-four. Two independent reviewers graded the same forty questions separately first, to check whether the rubric itself even meant the same thing to two different people, and they agreed on thirty-one of them, seventy-eight percent, a number Talitha wrote in her own notes as "not great, but real."

Run the same three weeks again, with retrieval accuracy checked first, on real question patterns, before anyone ever looks at what the model says about them. The missed premium shows up in week one, inside a review meeting, not inside Theron's account four days after he already moved the money. Ask Marigold holds for two more weeks while the team fixes how it searches custom categories. It ships clean.

What I'd want to tell myself, if I could go back to that planning meeting: a test written by the same three people who built the feature will always pass, because it can only ever ask the questions they already thought of.

ORDER, the five checks Marigold skipped

Not a way to make two kinds of "done" sound like the same idea with different names. ORDER is what forces you to name the second kind and rank it properly before anyone signs off on the first.

Five checks, run once on two features that were never going to fail the same way.

Same word, done. One of these mistakes reverses with a patch. The other one doesn't.

Retrieval gets checked first because nothing after it means anything if the search itself is wrong.

Three things worth naming directly, since this is where the real judgment sits. Talitha's team tried the cheap fix first: raise the golden set's pass bar from 94 to 99 percent. Rejected, because a harder version of a test the engineers wrote themselves still never asks a question shaped like Theron's. The AI-specific failure mode here is a silent retrieval miss: the model isn't wrong about what it read, it's wrong about what it never found, and a model reasoning confidently over a missing fact sounds exactly as sure of itself as one reasoning over a correct one. The guardrail is the retrieval-accuracy-first check, run before anyone grades what the model says about the data it pulled. And there's a real cost to doing this properly: building a representative eval set, a rubric, and a human-agreement baseline for Ask Marigold cost Talitha's team roughly two extra weeks before it could ship again. That's the trade Marigold is making on purpose: slower to ship a retrieval feature, in exchange for a done that can survive a real user's real question.

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

Same five letters, a completely different clinic. This time nobody's moving money. Somebody's deciding how fast to get a sick animal in front of a vet.

Furwell is a scheduling and advice app used by a small chain of veterinary clinics. Owners use it for two things: booking an appointment by hand, and asking a symptom-triage feature how urgent something is, "my dog's been vomiting since this morning, should I come in now or wait."

Booking sits where a normal checklist is still the right one. Triage sits where it isn't.

Run the same five letters. Outcome: appointment booking protects a correct calendar slot. Symptom-triage advice protects a defensible urgency call, on real symptom descriptions, that a vet could stand behind afterward. Reversibility: a booking bug means a double-booked slot, annoying, fixed with a phone call. A triage feature that quietly under-calls urgency on a real symptom, "come in within a week" when it should have said "come in tonight," can cost an animal's life, and there is no patch that reverses that. Dependency: the triage feature needs real symptom write-ups pulled from actual vet intake notes, a rubric a vet, not an engineer, agrees with, and a check that it escalates instead of guessing when the description is thin. Evidence: cheap first check, does the retrieval pull the pet's real medical history correctly, before anyone grades the urgency call it makes on top of that history. Rank: retrieval accuracy first, then the rubric-graded eval against real intake notes, then a vet's own sign-off on the escalation logic, then a small rollout, watched closely for two weeks before it goes wide.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: a save button's done is deterministic, checked once. A retrieval feature's done is a defensible bar on real questions, checked with a real eval set and a human-agreement baseline, and you check retrieval accuracy before anything else.
Cost: no budget this quarter to build a full eval pipeline. Ship the retrieval feature advisory-only, human-reviewed on anything above a small dollar or urgency threshold, until the real eval set exists, rather than shipping it unguarded to save two weeks.
The model got better, for real: say Ask Marigold's next retrieval upgrade actually closes the custom-category gap. The pass rate on the representative eval should climb toward the golden set's, but only if someone re-runs the real eval to check, instead of assuming a better model automatically means a better answer everywhere it's used.

Where people run it wrong.
They let the team that built a retrieval feature also write its only test, since nobody else has time.
They treat one blended pass rate as proof, instead of checking whether two people even agree on what counts as a pass.
They test a CRUD feature and a retrieval feature against the same checklist, because it's always been one checklist before.

How to use it live. Ask one question before trusting any "it passed its eval" claim: "who wrote the test, and would a real user's actual question have made it in?" That question alone usually tells you whether a pass rate is measuring the feature or just measuring its own builders.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0