Review and correct

You need: the reviewer dataset role or higher.

Review and correction are separate jobs on purpose — read Review and correction for why.

Setting up a review queue

1. Open the dataset's review page and configure a queue.
2. Pick a sort strategy — the order objects are presented in. Confidence-based orderings put the model's least certain output first, which is where reviewer time changes the most.
3. Create the queue. It is a snapshot: a fixed list in a fixed order, so it has a definite end and does not reshuffle underneath you.

Reviewing

Work through the queue giving each object a verdict. R rejects in review mode.

Rejected objects resolve as one of:

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 a real answer

Recording that a human looked and chose not to change something is different from nobody having looked. That difference is what lets you state the annotation quality of a dataset later.

Setup → sort strategy → session → one fixed, one wont_fix → queue progress.

Correcting

The correction pass is where rejected objects get fixed, using the same drawing and editing tools as annotation. Model output stays separable from human edits throughout, so a corrected annotation knows it was corrected and by whom.

Correction queue → fix a bad contour → model output vs. human edit distinguishable.

Next: Train a model on what you just corrected.