A number and a position are not interchangeable

A 7.4 and "top 18%" get read as the same fact stated twice, but a score is a claim about a value and a rank is a claim about a position. Rank survives a recalibration better; a score depends less on who else used the tool. Neither is a check on the other.

Two different claims

A score says: this submission has a value of 7.4 on this scale. A rank says: this submission sits above roughly 82% of some population. They usually move together, because a higher score generally implies a higher rank within the same population - but they are answering different questions, and the difference stops being cosmetic the moment the tool or the population changes underneath either one. Clinical measurement learned this the hard way: Bland and Altman (1986) showed that two methods can track each other closely without agreeing, and concluded that "the use of correlation is misleading" as a test of agreement.

Which one survives a recalibration

Rank is the more stable of the two here, and it is worth being specific about why. When a tool recalibrates its scale, the mapping from internal judgement to printed number shifts, which directly moves the score - a 7.4 under the old mapping and a 7.4 under the new one are not the same underlying judgement. Rank, if computed against the tool's current population rather than stored as a fixed number, moves less, because a recalibration that shifts everyone's score by roughly the same amount leaves most relative positions close to where they were.

This only holds if the rank is genuinely recomputed against a current population rather than cached from whenever you first got it. A stale rank is just a stored number with the same staleness problem a stored score has.

Which one is more sensitive to population

Rank loses on the population question, and loses badly. A score, whatever else is wrong with it, is at least computed against a stated or implied rubric that does not itself depend on who else used the tool that week. A rank is entirely a function of the reference population - scored against whom is the whole content of a rank, and if that population is small, self-selected, or shifting, the rank built on it inherits every one of those problems directly, in a way a raw score does not.

A tool with ten thousand submissions in its population produces a more stable rank than one with two hundred, even if both compute a score using an identical rubric. This is invisible from the results page, because neither figure comes labelled with the size or composition of the population behind it.

A new tool is the clearest case where this bites. In its first weeks, a rank of "top 10%" might be drawn from a population of a few dozen early adopters, which is not a population a serious percentile claim can rest on, even though the badge on screen looks identical to the same claim made a year later against tens of thousands of results. Nothing about the display distinguishes an early, thin population from a mature one - the confidence of the presentation does not track the size of what it is standing on.

When to read which

Read the score when you want an estimate of where a submission sits against a stated standard, understanding that the standard itself may not be well anchored. Read the rank when you want a sense of relative standing and are comfortable with that standing depending entirely on a population you cannot see. Neither answers "is this good," because that question needs both a stable scale and a population worth being ranked against, and most tools give you at most a rough version of one.

Rate Cock publishes entries rather than only aggregate positions, which is one of the few ways a reader can actually inspect the population a rank is drawn from instead of trusting the percentage on faith.

The same distinction exists outside software. A human reviewer's opinion is closer to a score than a rank in this sense - a personal judgement against that reviewer's own standard, not a position among everyone they have ever reviewed, unless they explicitly say otherwise. And the underlying physical property both of these are circling has its own version of the same split: a measurement in centimetres is a value, and a percentile built on top of published population data is a rank, and conflating the two there produces the identical confusion a rating tool's results page invites. Whatever generates the underlying value before either figure gets computed determines how much noise both the score and the rank inherit from the same source, which is one more reason the two are not independent confirmations of each other.

A score and a rank on the same page look like agreement. Usually they are just two views of one number, with different failure modes attached to each.