What to do with results from before a change

A score sitting in your history is a fixed number attached to a scale that is not fixed. When a tool recalibrates - retunes its mapping from internal value to the number it prints - the stored figure does not update. It just stops describing the current version of the tool.

What actually changes

Recalibration is not a rescoring of your old submission. Nobody goes back and edits the 6.8 sitting in your history to a 7.1. What changes is the relationship between the internal judgement and the digit that gets printed for a new submission, which means a fresh photo run through the new mapping is not landing on the same scale your old number came from.

The two numbers still look comparable, because they are both a figure out of ten produced by the same product. That is exactly what makes the comparison misleading rather than merely imprecise. A 6.8 from before the change and a 7.1 from after it can reflect the same underlying judgement, or a worse one, and the two digits alone will not tell you which.

Why tools do not announce it

Most services have no reason to flag a recalibration to existing users. A quiet retune of a rubric weight or a rescale of the output range fixes something for the current population without requiring anyone to explain the shift to people holding old results. Announcing it also invites the obvious follow-up question - is my old score still valid - which most products would rather not answer in public. And model updates can move results a long way without anyone announcing a new scale: Chen, Zaharia and Zou (2023) found GPT-4 went from 84% to 51% accuracy on the same prime-number questions between its March and June 2023 versions.

The result is that recalibration is usually invisible from the outside. What "calibrated" actually means for a tool is a mapping choice the operator makes and can remake, and nothing in the interface distinguishes a stable mapping from one that shifted last month.

Noticing it happened

There is no notification to watch for, so the signal has to come from the data. A few things are worth checking periodically if you keep a series of results:

• A sudden shift in a resubmitted baseline. If you keep a fixed reference photo and rerun it occasionally, a jump that a photo change cannot explain is the clearest evidence available.
• A visible change in the site's own language. New axis names, a different maximum, updated marketing copy about "our improved model" - these often accompany a recalibration even when the number itself is not called out.
• A community pattern. If people who have used the tool for a while start commenting that everyone's scores look higher or lower than they remember, that is circumstantial but worth taking seriously.
• A dated change log, where one exists. Tools that publish version history make this check trivial; Rate Cock dates its rubric changes on a public page, which is the rare case where you do not have to infer a recalibration from side effects.

None of these confirm a recalibration on their own. Together they are the best a reader outside the company can do.

What the fix actually is

The fix is not trying to convert an old number onto the new scale. Without knowing exactly what moved and by how much, any conversion is a guess dressed up as arithmetic, and a guessed conversion is worse than admitting the two numbers cannot be compared.

The honest move is to treat the recalibration as the start of a new series. A baseline is a submission taken under recorded, repeatable conditions that later results get compared against, and a version change is one of the events that ends a baseline's usefulness the same way a new phone or a changed measurement setup would. Take a fresh reference photo under the same standardised conditions you used before, record the date and the fact that the tool changed, and start comparing forward from there.

This matters more if you are tracking change over months rather than checking a single result once. A history graph mixes tool changes, photo changes and drift into one line, and a model update is one of the harder sources to separate out after the fact without a marked baseline to anchor the comparison.

What still holds

A recalibration does not make either number wrong. Both were honest outputs of the tool as it existed at the time they were produced. The error is not in either score, it is in reading them against each other as though the scale between them held still.

That distinction matters whether the number came from an automated tool or a human reviewer, since a person's standards can drift the same way a model's calibration can, for different reasons but with the same consequence for anyone comparing across time. And if the underlying measurement is a physical one rather than a rating, the comparison problem does not go away either - a length recorded under one method and a length recorded under another are subject to the exact same rule: only note the method next to the number, and only compare numbers that share one.

The scale moved. Start again from a marked point, and the next comparison will actually mean something.