Skip to content

Review and correction

Review and correction are two different jobs, and IQUANA keeps them separate.

  • Review asks is this right? — a verdict on existing annotations.
  • Correction is the work of making it right — editing the ones that are not.

Splitting them means a reviewer can go through a hundred predictions quickly without switching into drawing mode for each one, and the corrections can be done in a focused pass afterwards, possibly by someone else.

Queues are snapshots

A review queue is a snapshot taken when the queue is created — a fixed list of objects, in a fixed order.

This matters more than it sounds. A live query would reshuffle under the reviewer as annotations changed around them, making "how far through am I?" unanswerable and making two reviewers' work impossible to compare. A snapshot is a stable unit of work with a definite end.

Ordering

Queues are built with a sort strategy — the order in which objects are presented. The strategies are a registry rather than a fixed list, which is what allows active-learning orderings: surface the objects the model was least confident about first, so reviewer time goes where it changes the most.

Verdicts

A reviewed object ends in one of two resolutions:

prediction model output in a review queue a snapshot: fixed list and order approved looked right rejected something was wrong fixed corrected wont_fix left deliberately queued approve reject revoke
The states a prediction moves through. Both terminal resolutions are recorded outcomes - wont_fix is an answer, not an absence of one.
Resolution Meaning
fixed it was wrong, and it has been corrected
wont_fix it was wrong, and it is being left that way deliberately

wont_fix is not a shrug. Recording that a human looked at something and chose not to change it is different from nobody having looked, and the difference matters when you later ask what the annotation quality of a dataset actually is.

Provenance

Model output stays separable from human edits throughout. A corrected annotation knows it was corrected, and by whom — see Roles and permissions for how attribution works across a team.

Review setup → pick a sort strategy → session → one fixed, one wont_fix → progress through the queue.
Correction queue → fix a bad contour → model output vs. human edit staying distinguishable.