← Back to concept explainers

Concept Fable

Distribution shift

Conclusion first:Distribution shift means the data distribution a model faces during training, validation or historical evaluation differs from the real input distribution after deployment. It can make offline scores look good while online performance degrades. Common shifts for enterprise Agents include changes in customer segments, industry language, task types, tool return formats, channels and policies. Governance approaches include segmented monitoring, pre-launch scenario coverage, small-traffic trial runs, drift detection and continuously updated evaluation sets.

Distribution Shift · Evaluation

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The foot-measuring ruler arrived at another dock

There was a shoe shop in Lugang that initially made boots only for local fishermen. The shopkeeper had measured many feet and knew how high the instep was, how thick the sole should be, and how much margin to leave for the rainy season.

The shoe shop later received a large order from an out-of-town dock. The manager still worked from the old size chart, because past repairs had been few and the ledger proved the sizes reliable. After the first batch of boots was delivered, returns kept coming. Dock workers carried salt sacks year-round, so their feet were wider; they walked on stone paving, not mudflats; and they wore boots for longer periods.

The shopkeeper's naive fix was to enlarge all boots by one size. Returns decreased a little, but new problems appeared: some people got blisters, some tripped, and some found the soles too soft. The assistant then suggested looking only at the single most common return reason. But each dock had different problems, and moving a patch from one place to another created new errors instead.

The old shoemaker took people to the new pier to measure feet, observe roads, and ask about working hours, and compared these records side by side with the original Lukang data. For the first time, they saw that the two groups of customers did not come from the same world.

The shoe shop re-divided its size chart: fishermen, salt workers and ship repairers each had their own samples and acceptance rules. Orders from new places had to run a small trial batch first before deciding whether to continue using the old machines. Later, whenever the shoe shop arrived at a new dock, the first thing was no longer to start work as usual, but to confirm whether the people, roads, weather and usage patterns of this place still matched the old ledger.

The shopkeeper first blamed the craftsmen for being lazy, then had everyone sew the stitches more densely. The boots became sturdier, but still pinched, because the problem was not the craftsmanship but that the feet of the old town had been treated as the feet of every port.

He took a ruler and stayed out of town for seven days, measuring feet, observing roads, asking about working hours, and taking apart every returned boot one by one. The new size chart changed only a few inches, yet it overturned many assumptions.

The shoe shop later still kept the old ledger, but tried on every new dock first. Past good results can only explain the past; they will not guarantee the future in an unfamiliar place. The shopkeeper finally understood that a ruler being accurate in one place does not mean it is still accurate when the people and setting change.

Reveal

This story is about: distribution shift

Distribution shift means the data distribution a model faces during training, validation or historical evaluation differs from the real input distribution after deployment. It can make offline scores look good while online performance degrades. Common shifts for enterprise Agents include changes in customer segments, industry language, task types, tool return formats, channels and policies. Governance approaches include segmented monitoring, pre-launch scenario coverage, small-traffic trial runs, drift detection and continuously updated evaluation sets.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Lukang fishermen: the populations and scenarios covered by training data
  • Out-of-town dock workers: new input distribution after deployment
  • Old size chart: model patterns learned from historical data
  • Uniformly enlarge by one size: a rough patch that cannot resolve distribution differences
  • Go to the new dock to measure feet and survey the road: collect production samples and analyze subgroup differences
  • Small-batch trial wear: canary, pilot or shadow evaluation
  • Segmented size chart: build evaluation and monitoring by scenario, industry and user group
  • Final insight: Distribution Shift means historical performance no longer automatically represents future performance

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

generalizationrobustnessdomain adaptationtest set
京ICP备2026014272号