Handling an outlier in a series

Keep a wild retake unless your log shows a documented change in conditions on that attempt. Four results cluster, a fifth sits a point or more away, and the temptation is to drop it as a bad shot. That instinct is right about as often as it is wrong, and the difference is checkable rather than a guess.

The two things an outlier can be

The first possibility is that something in your conditions actually slipped on that one attempt: the phone tilted a little more than the others, the light changed because a cloud moved, the crop was eyeballed instead of matched to the rest of the set. In that case the outlier is real in the sense that the input genuinely differed, and the score is doing its job by reflecting a real difference in what it was given.

The second possibility is that nothing about the conditions changed and the tool itself produced an unusual read - a rare miss on one axis, a processing hiccup, ordinary noise landing at the tail of its distribution instead of near the centre. In that case the outlier is not evidence of anything about your submission; it is the tool's own variance showing up, and it belongs in the average like any other draw.

Check the log, not your memory

The only reliable way to tell these apart is to have written the conditions down before you took the shot, not to reconstruct them afterward from a feeling that something seemed off. Logging distance, angle, light and crop for each attempt is what turns "that one felt different" into "that one was different, here is what changed" - and it is the only version of that sentence worth acting on.

If the log shows a real difference on the outlier attempt, it is fair to treat that result separately from the rest of the series - it measured a different condition, not the same one five times, so it does not belong averaged in with attempts that held the protocol constant. If the log shows nothing different, the honest move is to keep the result in the set. Deleting an inconvenient number because it does not match the story you expected is not cleaning the data, it is choosing the story. Statisticians take the same line: the NIST/SEMATECH e-Handbook of Statistical Methods notes that outliers "may be due to random variation" and that "we typically do not want to simply delete the outlying observation".

What to do with it once you have decided

A kept outlier still belongs in a series that gets summarised the ordinary way, as a centre and a spread rather than a single figure. The median is naturally resistant to exactly this kind of single wild value, which is part of why it is the right summary statistic for a short series in the first place, though the case for the median on its own terms is a separate argument. A wide range that includes the outlier is a more honest report of what the tool produced this session than a narrower range built by quietly excluding the one result that did not fit.

This same discipline shows up wherever repeat measurement matters. A single unusual reading from a tape measurement gets the same treatment - check the method, not the number, before deciding it does not count - and how a measurement's own conditions get logged and reproduced is built around exactly this kind of accountability. A single unusually generous or harsh session from a human reviewer raises the same question in a different form; what makes one reviewer's read land outside their own usual range is worth a glance at their stated approach before assuming the read itself was wrong. And on the tool side, an outlier can also originate somewhere you cannot log at all - variance that comes from the model's own inference rather than from your setup is a case the log will correctly show as "nothing changed," which is itself the signal to keep the number rather than discard it.

The rule is small and worth keeping close: an outlier earns exclusion by pointing at a documented cause, not by looking inconvenient. Everything else, kept or not, still needs the wider treatment a single session gives it - a short series summarised properly rather than a single number quoted as if it were the whole story, on Rate Cock or anywhere else that logs more than one attempt.