Scores
Where a score gets rounded, and what that hides
A rounded score can move a whole point on a change too small to see; rounding is a threshold, and thresholds create false jumps.
Guides on Scores: How much weight a rating tool result deserves, Every part of a rating result, and what each one is for, What can and cannot be compared, and how
Rating tools round a more precise internal value to one decimal or a whole badge number before you see it. The rounding is harmless; what it hides is that a tiny movement near a boundary can land on either side of it, and only one side gets celebrated as a change.
The 6.96 to 7.04 case
Say a tool's internal value moves from 6.96 to 7.04 between two results - a shift of eight hundredths, well inside the noise of an ordinary retake, smaller than the gap you would expect between two photos taken thirty seconds apart. Displayed at one decimal, or rounded further to a whole badge number, that becomes a jump from 7.0 to 7.0, or in some display schemes, a crossing from "6" territory into "7" territory. A tool that shows a labelled tier - "good" starting at 7 - turns an eight-hundredths shift into a tier change, and a tier change reads as an event. Nothing meaningfully happened in the underlying judgement. The number simply crossed a line drawn at a round figure, and the line is where the drama comes from, not the change itself. Rounding conventions are standardised - NIST's Guide for the Use of the SI (SP 811) even specifies that an exact trailing 5 rounds to leave an even digit - but no convention removes the boundary itself. Metrology also keeps displayed digits apart from real sensitivity: the International Vocabulary of Metrology defines resolution as the "smallest change in a quantity being measured that causes a perceptible change" in the indication.
Why this matters more at some boundaries than others
Any score sitting close to a rounding boundary is one ordinary retake away from appearing to move, in either direction, for no real reason. A score that sits at 7.5, well clear of the nearest boundary, is comparatively stable in its displayed form even if the underlying value wobbles by the usual amount between attempts. A score sitting at 6.98 is one normal fluctuation from crossing into "7," and the crossing will look, to anyone reading only the displayed number, like the exact kind of improvement they were hoping a change actually produced. The honest read is to notice when your result sits near a boundary and discount a crossing accordingly - a jump from 6.9 to 7.0 is not more meaningful than a movement from 6.9 to 6.95 that never crosses anything, even though only one of the two looks like news.
This is a separate issue from what the decimal itself represents once it is on the page - that question, about what the second digit in a 7.4 is actually claiming, is worth reading on its own, since rounding and precision are related but not the same problem. It is also worth reading against how a half-point change should be interpreted generally, since a boundary crossing is one specific, more visually dramatic version of that broader question.
Tools that show a breakdown rather than a single rounded total give you a way around this - Rate Cock's per-axis figures make a boundary crossing on the total less likely to be mistaken for a real event, because you can check whether the axes actually moved together or whether one rounded figure alone did the crossing. The same caution applies to any number reported with a fixed threshold, including a measurement's stated tolerance and a model's own accuracy figure - a boundary is a display choice, not evidence that something on either side of it is different in kind. None of it applies to a human reviewer's response, which was never rounded to begin with.