Do timestamp errors average out?
No — and that single fact reorganizes the whole subject.
Most measurement thinking assumes you can measure again.
The archer shoots ten arrows and the group tells you something the single arrow could not. The metrologist repeats a reading and quotes a mean. Nearly every technique on this site — averaging, precision, the whole idea of a distribution — assumes repetition is available.
Timestamping a network packet is not like that.
The packet passes the timestamping point only once. The timestamp is taken once. There is no second observation to average with, no group to draw, and no opportunity to notice from the spread that something was off. You get one number and you keep it forever.
What that changes
Precision stops being a statistical comfort. A system whose timestamps scatter by ±50 ns does not give you a nice tight distribution around the truth; it gives you one sample from that distribution, and you will never know which one you got. The spread that would have averaged away is, for this single event, simply your error.
Bias and scatter become equally bad. Elsewhere the distinction matters enormously — bias survives averaging and scatter does not. With one shot there is no averaging, so both land in the answer with equal weight. The careful separation of trueness from precision is still the right way to diagnose a system, but it stops being a way to escape one of them.
The error is unfalsifiable after the fact. Nothing in the record says which timestamps were the good ones. A packet recorded 80 ns late looks exactly like one recorded on time, forever, and no amount of later analysis separates them — which is why a regulator asking about a trade from last quarter is asking a question the record cannot answer.
Which is why the work happens beforehand
Everything you can do about single-shot error has to be done before the shot.
This is the sharp edge of will averaging longer help? That page says averaging buys precision and never buys trueness. This one says that when you only get one measurement, you do not even get the precision — so the two failures that page keeps carefully apart arrive together, and both of them arrive in your answer.
You cannot average away a systematic offset in a timestamp you have already taken — but you can have measured your feedline last month, so the offset was never there. You cannot improve the precision of a record you already have, but you can have chosen a clock whose scatter was small before you ever wrote to it.
This is the underlying reason the questions on this site are shaped as things to ask before buying. In a repeated-measurement world, characterizing the instrument afterwards is a reasonable plan. Here, afterwards is too late by construction.
The archer’s target is still the right picture for understanding the kinds of error — it is used throughout the metrology pages. Just remember that in this application you are looking at a target with exactly one arrow in it, and no way to tell whether it was a representative one.