Annotation, in preview.
A walkthrough of the annotation surfaces — offline-durability studio, pixel-mask with SAM2 assist, video action timelines, voice comments, structured bulk-edit, calibration-aware cuboids, Quality Hub, the export gate, and adjudication auto-escalation. All counts read from a checked-in metrics snapshot.
Human Data OS · annotation
Reviewed work, with its proof attached.
Image, video, audio, and text share one surface. Reviewer identity, quality score, and export state travel with the record, so the dataset you ship can answer who labeled it and who approved it. The numbers below are illustrative seed data; in a pilot they read from your scoped review metrics.
Mark appeal language and route policy risk to review.
00:14
Connection drops
Network partition detected · work queues to IndexedDB.
00:41
Offline durability
Operator keeps labeling · 37 ops queued locally.
01:03
Reconnect + replay
Queue drains in order · CRDT resolves overlaps.
Studio durability
Offline queue. Reconnect replay.
When a connection drops, the editor queues every operation to local storage and replays them in order the moment the socket is back. No lost work, no duplicate labels.
- 00:00
Connected
Editor online · autosave to server every 10s.
- 00:14
Connection drops
Network partition detected · work queues to IndexedDB.
- 00:41
Offline durability
Operator keeps labeling · 37 ops queued locally.
- 01:03
Reconnect + replay
Queue drains in order · CRDT resolves overlaps.
Durability is backed by IndexedDB + a CRDT replay log on the Studio client.
Image · pixel-mask + SAM2 assist
Mask, correct, confirm.
SAM2 proposes a pixel-mask. The annotator accepts, corrects, or redraws — every stroke stays attached to the record.
Tool palette
- Polygon
- Brush
- SAM2 assist
- Eraser
SAM2 proposals come from the separately-deployed inference server; the editor never blocks while proposals compute.
Video · action recognition
Tracks, per action, per record.
Three tracks. Each segment is a labeled action in the clip. Scrubbing updates the spans — the review queue sees the same timeline the annotator drew.
Audio · threaded voice comments
Review by voice. Stays on the record.
Reviewers leave voice notes directly on a segment. The thread, the waveform, and the transcript all stay on the same task record instead of scattering across chat tools.
- Lead reviewer0:14
Missed the silence at 00:04 — re-run diarization.
- Annotator0:21
Agreed. Updated boundary + attached transcript.
- QA0:09
Looks clean now. Approving the segment.
Structured · bulk edit + taxonomy
Tabular work, schema-aware.
Apply a taxonomy change across a selection. The schema editor and the bulk editor share the same source of truth — no downstream re-imports.
Taxonomy editor
- Transactions
- returns
- disputes
- refunds
- chargebacks
| id | category | priority | state |
|---|---|---|---|
| txn-00481 | returns | p2 | queued |
| txn-00482 | returns | p2 | queued |
| txn-00483 | disputes | p1 | queued |
| txn-00484 | returns | p2 | queued |
| txn-00485 | refunds | p3 | queued |
3D · calibration-aware cuboid
Cuboids that carry their calibration.
LiDAR and depth annotations stay attached to the sensor frame, extrinsics, and time sync they were drawn against. Export-ready without a recalibration pass.
Calibration envelope
- Sensor frame
- lidar_top
- Extrinsics RMSE
- 0.014 m
- Intrinsics version
- 2026.03
- Time sync
- ±2.8 ms
Quality Hub · preview
IAA. Drift. Reopen rate. One hub.
Trend line, drift heatmap, reopen sparkline, and the active rule set live on one page — the evidence that says this corpus is good enough to train on, before anyone exports it. See how the rules carry through to a signed dataset on Human Data OS.
IAA median · 30d
0.84Target 0.75 · trend up
Drift heatmap · per annotator / per week
Cells above 0.05 force adjudication automatically.
Reopen rate · 30d
1.8%Ceiling 2.0% · below target.
Active rules
IAA floor
Blocks export when IAA median < 0.80 across the gold set.
Reopen-rate ceiling
Auto-escalates when reopen rate > 3% for any annotator.
Drift guard
Forces adjudication when drift vs baseline > 0.05.
Export gate
Weak data holds before export.
When the EU AI Act provenance rules take effect in August 2026, every high-risk training set has to prove its quality. Four checks run before an export is released. If any one fails, the export holds until the team clears the finding — and the passing run carries a record you can put in front of an examiner.
- IAA floor0.84 / 0.80
- Reopen rate1.8% / 2.0%
- Gold-set coverage82% / 95%
- Adjudication backlog0 / 0
Seed data: 26 of 30 export jobs passed the gate. In a pilot, this reads from your scoped review metrics.
Adjudication · auto-escalation
From disagreement to decision.
Threshold rules watch the queue. When agreement drops or reopen rate spikes, records auto-route to adjudication with the evidence already attached.
- 01
Disagreement detected
Two labelers, same record, conflicting spans.
- 02
Threshold rule evaluates
Reopen-rate + drift checks against live thresholds.
- 03
Auto-escalate to adjudicator
Rule fires — record routed to the adjudication queue.
- 04
Decision + audit
Adjudicator resolves; evidence + timestamps stored on record.
Run it on your own work
Read-only here. Hands-on in a pilot.
These are checked-in platform previews. In a pilot, point the same workspace at your data: the figures read from scoped workspace metrics, and the export gate produces a dataset record showing who labeled it and who approved it.
Continue through the review path.
Open the Annotation product page for the full story, or inspect the Quality Control Hub evidence in the preview.