Understanding Confidence Scores
What the numbers mean
Each screened paper gets a confidence score from 0-100%. Here's how to read them:
| Score | What to do |
|---|---|
| 90-100% | Usually correct. Accept in bulk to save time. |
| 70-89% | Probably correct. Give a quick glance. |
| 50-69% | Uncertain. Review these carefully. |
| Below 50% | Unreliable. Always review manually. |
Recommended workflow
1. Sort by confidence (ascending), tackling the hardest decisions first
2. Review the uncertain pile (below 70%), since this is where your time matters most
3. Bulk-accept high confidence decisions, accepting all above 90% to save time
4. Spot-check a few high-confidence decisions to validate
Calibration run
Before screening your full project, you can run a calibration set (20 papers). You screen them manually, then the AI screens the same 20 blindly. Compare the results to see if the AI matches your judgment before committing.