is younger than most people assume and is frequently asked to do work it was never designed for.
It became routine at most journals only in the second half of the twentieth century.
Its function is filtering and improvement, not .
A reviewer reads a description of an experiment and cannot know whether the experiment took place as described.
Expecting review to detect fraud is therefore a category error that damages trust when the inevitable case appears.
Studies of the process itself report low agreement between reviewers on the same .
Agreement on rejection is somewhat higher than on acceptance, which is what one would expect from a filter with a high threshold.
has been trialled extensively and produces smaller effects than advocates on either side predicted.
Author identity is often guessable from the , particularly in a small field.
The most consequential failure is not bias but delay.
A year between submission and publication was tolerable when the alternative was nothing.
servers have removed that condition, and the community now reads the preprint and cites the journal version.
Review has thus become a layer applied after the knowledge has already circulated.
That is a defensible role and it is not the role the system describes for itself.
Reviewing is also unpaid, and concentrated on a minority of the community.
Journals that publish reviewer reports alongside accepted papers have found that quality rises and volunteers do not disappear.
Attribution again does most of the work, which is a recurring result across every part of academic labour.
None of these reforms addresses the question of who decides what is worth reviewing at all, which remains with editors and is discussed least.