Skip to main content
A scan is one run of one engine against one dataset (or branch). It produces results, artifacts, and a severity. This page covers the launch, the lifecycle, and how to triage the output. For what each engine actually detects, see Engines.

​
Launching a scan

From a dataset detail page, click New scan.
1

Pick an engine

The picker lists every engine compatible with the dataset type. Image datasets surface mislabel, mislabel_broad, poisoning, bias_shortcut, mask_quality, and yolo. Text datasets show text_analysis. 3D / NIfTI datasets show mislabel_3d.
2

Pick a branch (optional)

If the dataset has branches, the scan defaults to the main view. Switch to a branch to scan just that manifest.
3

Tune parameters

Every engine accepts the same top‑level controls:
  • Use GPU, when a GPU worker is available.
  • Sample cap, limit the number of samples for fast previews.
  • Advanced parameters, engine‑specific knobs (thresholds, sensitivity, noise fraction, consensus level). The defaults work for most datasets.
4

Launch (or queue multiple)

Click Run to start. You can queue multiple engines in one submission; each one becomes its own scan row and runs in parallel as worker capacity allows.
Queue mislabel, mislabel_broad, and bias_shortcut together on a fresh dataset. The combination gives you label quality, OOD contamination, and shortcut risk in one pass.

​
The scan lifecycle

Scans move through four states, with live progress streamed over a WebSocket.
If your network blocks WebSockets, progress falls back to polling every few seconds. No configuration required.

​
Cancelling a running scan

From the scan detail page, click Cancel. Partial results are discarded and the scan is marked error with reason="cancelled". Audit log captures who cancelled and when.

​
Scan list (/scans)

A global table of every scan across the workspace.
  • Filter by project, dataset, engine, status, severity, time range, or user.
  • Sort by any column.
  • Bulk re‑run, cancel, delete.
  • Export the filtered view to CSV for offline review.

​
Scan detail (/scans/:id)

The scan page is where most triage happens.
The threshold graph is interactive: dragging the cutoff line re‑filters the issue table immediately. Use it to find a sensitivity setting you trust before triggering a healing run.

​
Result detail (/results/:id)

When you click a row in the issue table, you land on the per‑result page.
  • Full‑size preview of the sample.
  • Every score and per‑signal breakdown.
  • Status dropdown: pending, confirmed, dismissed, fixed. Confirmed issues become weighted signal for future scans; dismissed ones lose weight.
  • Comments with @mention to loop in a reviewer.
  • Neighbors, the nearest samples in embedding space for context.
  • Similar issues, other flagged items with the same signature.

​
Bulk triage

From the issue table on the scan detail page:
  • Select rows with checkboxes (Shift+click for range select).
  • Confirm / Dismiss / Mark for healing with the bulk‑action bar.
  • Export selection as CSV for offline review.
  • Send to Jira as one issue per row, or one rolled‑up issue with every row as an attachment. Snapshots travel with the issue.
Open the two scans in adjacent browser tabs, or use the Compare button on the scan list. The compare view shows threshold differences, distribution shifts, and a diff of which samples were newly flagged or newly cleared. Useful when iterating on parameters.
From the scan detail page, Re‑run opens the new‑scan modal pre‑filled with the previous parameters. Adjust what you want and submit. The new scan stays linked to its predecessor in the dataset’s scan history.
Bulk‑select the rows you want to ship, Send to Jira with “one rolled‑up issue”, and assign the resulting ticket. Engineers open the ticket and click straight back to the scan in Blindsight for context.

​
See also

Engines

What each engine detects and how to interpret its output.

Healing

Turn a scan’s findings into a cured copy of the dataset.

Scheduled scans

Run scans automatically on a cron.

Webhooks

Push scan events to your downstream systems.