Reviewing and Correcting Low Confidence Reads

Even with the cleanest slate cards, some reads will be wrong. Slate Tracker collects every uncertain read into a Review queue so you can confirm, correct, or reject in batch.

Where the queue lives

Two places show the same queue:

• Local dashboard: http://localhost:7745/review
• Cloud dashboard: dashboard.slate-tracker.com/review

The local dashboard is faster and works offline. The cloud dashboard is handy if you want a colleague to do the QC pass remotely.

What lands in the queue

An entry shows up in the queue if any of the following are true:

• Confidence below 80 on any required field.
• Required field missing (scene or take).
• Slate text matches multiple cues on the desk (only when desk integration is on).
• Manual flag (you can right click any entry and pick Send to review).

You can change the 80 threshold in config.yaml:

review:
  min_confidence: 80
  required_fields: ["scene", "take"]

Working through the queue

Each entry shows:

1. The image (click to view full size).
2. The text the recognition engine read, per field.
3. The confidence score per field, colour coded amber for warning and red for critical.
4. A row of action buttons.

The action buttons are:

• Accept: keep the read as is. The entry is removed from the queue.
• Edit: correct any field. Saving will rename and move the file on disk to match the new values.
• Re-run vision: ask the recognition engine to read the slate again. Useful if the first read tripped on glare. Counts against your monthly quota.
• Reject: drop the image from the library. The file is moved to .trash/ in your output root.

Keyboard shortcuts

The review queue has shortcuts so you can fly through it:

Bulk operations

Hold Shift and click to select a range. The action bar at the bottom of the page applies to every selected entry.

The most useful bulk action is Re-tag scene. If all entries in a range should have the same scene number (the slate marker was wrong on every take), you can fix them in one go.

Corrections improve future reads

Every correction is recorded with the original image. Next time the agent sees a similar slate (same handwriting, same field layout), it uses your correction history to bias the read. The improvement is gradual but real, especially across long productions.

You can review and clear your correction history in Cloud dashboard, Settings, Vision history.

What if I delete the file by accident?

The .trash/ folder in your output root keeps every rejected or overwritten image for 30 days. Drag a file out of there and back into the watch folder to re-process it.