How do you handle a migration that improves most cases and regresses a critical one?

How to answer this, stage by stage

Seven moves. Name the leading signal before the story, or it sounds like Iskra just got lucky finding the regression.

Let's learn

Here's what happens when a model gets dramatically better at almost everything, and quietly worse at exactly the one thing that matters most.

Say a company sells a tool that drafts the first version of a radiology report. A chest X-ray comes in, the model reads it and writes a draft: what it sees, what's normal, what needs a second look. A radiologist reads the draft, fixes it, and signs their name to the version that goes in the patient's chart.

The model in production catches about 84 percent of what a final signed report agrees with, averaged across every kind of finding on the sheet: pneumonia, fluid around the lung, an enlarged heart, and more. Every candidate gets checked against the same fixed set, every single release: 4,000 already-signed chest X-rays, read months ago by real radiologists, used as the answer key.

A new candidate, RX-5, pushes that number to 93 percent. Nine points, in one release. It writes cleaner language, catches more of the common findings, hedges less. Everyone who sees the demo likes it.

Here's the turn. The nine points are not the story. Buried in the same test, one row moved the other way: pneumothorax, a collapsed lung, the kind of finding that can turn fatal within hours if nobody acts. Pneumothorax recall dropped from 91 percent to 67 percent. Nobody who only looked at the top number would ever have seen it, because pneumothorax is 24 rows out of 4,000, about six tenths of one percent of the set. Even a total wipeout on that row would only have cost the aggregate about half a point. The 23-point gain everywhere else buries it completely.

This is a quiet case of silent degradation: a regression that trips no alarm, because the one dashboard built to catch trouble is exactly the one number built to average it away. The guardrail is the per-finding floor itself, checked before ship, not a promise that it can never happen again.

At its worst, this ships quietly. To find out whether it actually would have, the team ran RX-5 in shadow mode: it drafted real incoming scans at Cavendish Medical for three weeks, silently, next to the model that was actually live and actually seen by radiologists. Five real pneumothorax cases came through in that window.

The clipboard was telling the truth about the average. It just wasn't the truth about row fourteen.

What I'd leave alone. The other twenty-some finding types on that sheet, cardiomegaly, mild atelectasis, degenerative changes, don't need their own gated floor. None of them are time-critical. A slightly worse first draft there costs a radiologist a few more seconds of editing. Gating all thirty rows the same way as pneumothorax would slow down every future migration for a risk that isn't actually there.

The lesson. One number was never going to be enough, because "better" and "better everywhere it matters most" are not the same claim, and a dashboard built around one blended score can only ever answer the first one.

Now here is the same thing as a story

Read this version when you've got the extra few minutes. The short one above is what you'd say in an interview. This is why it's true.

Iskra Kendrick has run the report-drafting product at Pellard Health for three years, and in the first year, every migration got the same treatment: a whole afternoon in a conference room, the team walking a candidate model's score finding by finding, all thirty rows, before anyone signed off on a release.

Then the team built a dashboard. One number, refreshed every night: the aggregate agreement score, how closely a candidate's draft matched what a radiologist actually signed, averaged across every finding type at once. A migration that used to cost an afternoon became a five-minute glance at a green number on a screen. For a year and a half, every candidate that cleared the bar really was better everywhere. The dashboard was never wrong.

The first few releases, Iskra still opened the full thirty-row sheet anyway, out of habit, even when the top number looked fine. After a dozen clean releases in a row, she started only opening it when the top number looked off. By release nineteen, she wasn't opening it at all. Twelve people on the team had shipped the last six candidates off the dashboard number alone, and every one of them had gone fine.

Release twenty was RX-5, due to replace the live model at Cavendish Medical that coming Monday, the biggest hospital account on Pellard's books. The dashboard read 93 percent, up nine points from what was live. Friday afternoon, with nothing wrong on the screen and a full weekend ahead of her, Iskra opened the thirty-row sheet anyway. She couldn't later say exactly why. Boredom, maybe. A habit dying slower than she thought.

Row fourteen was red. Pneumothorax: 91 percent on the model that was live, 67 percent on RX-5. She read it twice, thinking it was a formatting error from the export. It wasn't. She pulled the six builds that led to RX-5 and watched the same story twice: the aggregate score climbing in a straight line, pneumothorax recall falling off a cliff starting at build four, the exact release where the team had changed the training mix to lean harder on the common findings.

Iskra considered just shipping RX-5 everywhere Monday as planned, pneumothorax included, since it was under one percent of scans and the aggregate case for it was overwhelming. She rejected that within the hour: rare isn't the same as cheap, and a tension pneumothorax missed on a night shift can kill someone before the next morning's read catches it. She also considered freezing the whole release, holding RX-5 back entirely until every row cleared, common findings included. She rejected that too: radiologists were already spending real time every shift cleaning up RX-4's clumsier drafts of pneumonia and effusion reports, and holding nine real points of improvement back for months, to fix one row, traded a smaller certain cost for a bigger one.

What she actually did was block Monday's cutover for one finding only, and spend the weekend setting up a shadow pilot to check her own read of the eval set: RX-5 would draft real incoming scans at Cavendish Medical for three weeks, silently, alongside the model that was actually live and actually seen by radiologists. Five real pneumothorax cases came through in that window. RX-4, still live, caught all five. RX-5's shadow drafts caught two. One of the three it missed was a large, tension pneumothorax, the kind that needs a chest tube within the hour, not the next morning.

I want to say the problem was the model got worse. It did, on one row. But that's not really the story. The dashboard never held thirty numbers in anyone's head. It held one, and that one number only had two settings: green, ship it, or red, stop and look. Nine straight points of green had flipped that switch for good, long before Friday.

A year earlier, when the dashboard first went up on the wall, Iskra had signed off on the ship rule herself: green means ship. It made sense that day. Every migration up to that point had genuinely moved every finding type together, because the training changes were broad enough that a gain anywhere showed up as a gain everywhere. Nobody had reason to think one dial could hide a needle moving the other way underneath it.

The fix took a weekend to design and eight weeks to actually clear: pneumothorax kept running through RX-4's older, still-91-percent check, gated separately, while everything else in RX-5 shipped that Monday as planned. A retrain, RX-5.1, oversampling pneumothorax-positive images specifically, brought recall back to 90 percent, and the finding cut over eight weeks later, on schedule and checked, not assumed. Four months after that, release twenty-six had the exact same shape: a nine-point aggregate gain and a small dip on a different critical finding, the cardiac silhouette obscured by a large effusion. This time the per-finding dashboard flagged it itself, in the Tuesday-morning review, in about twenty minutes. Nobody needed a lucky Friday.

One design put thirty numbers behind one door and trusted the door. The other keeps every door labeled, and the ones that can hurt someone locked separately.

And the thing I'd tell myself, back when we put that dashboard on the wall: a single number is only honest for as long as everything it's averaging moves together. The day that stops being true, and it always stops being true eventually, the number keeps looking fine right up until someone gets unlucky enough to notice.

LEAD, so the aggregate can't swallow the one row that matters

This sounds like a launch-readiness question. Underneath, it's still asking which number you trust when two of them disagree. That's LEAD, run on a migration gate instead of a dashboard.

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

Fenchurch Mutual runs an AI tool that scores incoming auto-insurance claims for likely fraud, routing the risky ones to a human investigator instead of paying them out automatically. Branwen Norwick, who owns that product, is looking at a migration candidate that lifts the overall fraud-catch rate from 61 to 74 percent, a genuinely large jump, and quietly drops the catch rate on staged commercial-fleet collisions, a rare, expensive pattern, from 82 to 45 percent.

Branwen set the same kind of floor: fleet-collision catch rate can't drop below 82 percent, checked monthly against the audited set. The new model ships for every other claim type, while the old model's fleet-collision check runs in parallel until a retrain clears the bar. It's a small trade against a big one: a handful of common claims wait an extra day or two for a human look, so that one expensive, rare pattern doesn't slide through unwatched while the fix ships.

Same shape as the chest X-ray story, a different claim file. The aggregate clock was never going to ring first.

Flashcards (click a card to flip it)

Check yourself Score: 0 / 0