Review settings
Project β Settings β Review.
Controls how reviewers receive tasks in the review stream and how their decisions (accept / fix & accept / reject) are recorded.
There is no on/off switch β every project has a review stream. Anyone who can review (project Reviewer, or Owner / Admin / Manager, the project creator, staff) can open it; projects that don't need review simply never send anyone there.
Visible to Owner, Admin, Manager.
Reviewer instructionsβ
- Instructions β HTML shown to reviewers
- Show before reviewing β open the instructions automatically when a reviewer enters the stream
Reviewing optionsβ
| Setting | Default | Effect |
|---|---|---|
| Show only manually assigned tasks | Off | On: reviewers see only tasks they were assigned to. Off: assigned tasks are served first; once the reviewer's assigned queue drains, unassigned tasks follow |
| Review workflow (enable review) | On | Off: no review stream, no Review button on the dashboard; annotations are final as submitted. Audio QC and crowd validation are independent and keep working |
| Show only finished tasks | Off | On: tasks not yet marked complete (is_labeled) are hidden |
| Tie-breaker mode | Off | On: only tasks whose annotators produced different results are served β the reviewer breaks the tie. Tasks where everyone agrees are skipped (the stream scans up to 200 candidates per request) |
| Task ordering | Sequential | sequential / random / agreement / model_confidence β see below |
| Conflicting annotations | Single ground truth | What happens when more than one annotation on a task is accepted β see Conflict policy |
| Task limit (%) | 100 | Share of project tasks included in the stream when ordering is Random β see below |
Regardless of these settings a task only enters the stream once it has
at least maximum_annotations non-cancelled annotations and at
least one of them is still unreviewed (or carries an open re-review
dispute). Partially-annotated tasks never surface.
Task orderingβ
| Order | Behaviour |
|---|---|
| Sequential (by Task ID) | Lowest task ID first |
| Random | Reshuffled on every request |
| By agreement (uncertain first) | Stored on the project; the stream currently serves these in task-ID order β use Tie-breaker mode to surface disagreements |
| By model confidence | Stored on the project; currently served in task-ID order |
QC priority. When the project's QC priority in review toggle (Settings β Audio QC) is on, tasks with an open review-queue item (QC failure, low agreement, outlier, eval failure, manual flag) are served first, ahead of whichever ordering is picked.
Task limitβ
Applies only when ordering is Random. A value below 100 caps the pool to that percentage of the project's tasks:
- The subset is picked once per project (seeded by project ID) and is the same for every reviewer; only the order reshuffles per request.
- Tasks with an open review-queue item are always included, even when they fall outside the sampled subset.
- 0 is a no-op β nothing auto-accepts; the full pool is served.
- The limit is ignored when reviewers open tasks directly from the Data Manager (selection or quick view).
Conflict policyβ
Governs accepting several annotations on the same task:
| Policy | Effect |
|---|---|
| Single ground truth (accept one) (default) | Accepting an annotation makes it the task's ground truth and clears ground truth on its siblings |
| Allow, but warn on conflict | Allowed; the review UI confirms when a second, differing annotation is accepted |
| Allow all (keep for agreement) | Every accepted annotation stands β use when the goal is inter-annotator agreement and a single label is produced later at export |
Reject optionsβ
- Reviewers must leave a comment on reject β enforced in the UI and by the server (a reject without a comment is refused).
Data Manager for reviewersβ
- Show the Data Manager to reviewers (default on)
- Show unused task-data columns to reviewers
- Show agreement to reviewers in the Data Manager
Accept / reject behaviourβ
A verdict sets review_state on the annotation (accepted,
fixed, rejected) plus reviewed_by / reviewed_at and the
optional comment.
| Decision | Effect |
|---|---|
| Accept | review_state = accepted; annotator credited in the contribution ledger |
| Fix & Accept | Reviewer's edits saved on the annotation; the annotator's original is snapshotted so the UI can show a "fixed in review" diff; review_state = fixed |
| Reject | review_state = rejected. The annotation stays in the database and is not re-served to annotators. The annotator can ask for a re-review (dispute); an open dispute puts the item back in the stream |
Reviewers can revisit reviewed items and change a verdict; the annotator is notified again and any credit that no longer applies is reversed. Annotators learn outcomes through a per-reviewer digest flushed when the reviewer's queue empties.
Managers can skip per-item review on wholesale-trusted projects: Project β Settings β Danger Zone β Accept All Pending Annotations accepts every unreviewed annotation and credits the annotators, after a typed confirmation showing the exact counts. Tasks with an open review-queue item are always skipped β see Accept all pending.
Assigning reviewersβ
Settings β Review β Assign reviewersβ¦ attaches reviewers to tasks in bulk:
- Pick reviewers from the project's members
- Choose the scope β all tasks, only tasks with unreviewed annotations, or specific task IDs
- Choose add (keep existing reviewers) or replace
Assignees get an in-app / email notification with a link to the review stream. Assignment only affects ordering (and hard-restricts when Show only manually assigned tasks is on); reviewers can always enter the stream without pre-assignment.
Review stream UXβ
- Task opens with its annotations; each shows
review_state, the reviewer's comment, any open dispute reason, and the fix diff - Accept / Fix + Accept (the label switches when the reviewer has edited) / Reject
- Keyboard:
Ctrl+Enteraccept,Ctrl+Spacereject (see Hotkeys) - After a verdict the next task loads; reviewed items can be revisited from the reviewed list
What's nextβ
- Review workflow overview β β the end-to-end reviewer journey
- Agreement metrics β β how agreement is scored
- Annotation settings β β the annotator-side companion settings