Metrics
Broken Evidence Rate: Calculation and Sources of False Positives
Broken evidence rate is the share of evidence links that fail, checked monthly in batches and reviewed by hand.
- Author
- Xingge
- Published
- 2026-11-09
- Updated
- 2026-11-09
- Level
- L2 Reproducible
- Version
- 1.0
Term details
- Name
- Broken Evidence Rate: Calculation and Sources of False Positives
- Type
- Metrics
- ID
- broken-evidence-rate-en
Broken Evidence Rate: Calculation and Sources of False Positives
Broken evidence rate measures the share of evidence entries whose links fail or whose pages are unavailable. It is calculated as failed entries divided by total evidence entries, inspected monthly, with false positives excluded by manual review before counting.
Definition and calculation standard
The denominator is the total number of registered source links, and the numerator is the number of entries that return an error status, redirect to an unrelated page, or require login to view. Timeouts and temporary failures are first marked for review and not counted directly.
| Item | Value |
|---|---|
| Metric name | Broken evidence rate |
| Formula | Failed entries divided by total evidence entries |
| Unit | Percentage |
| Recording frequency | Monthly |
| Collection method | Batch requests plus manual review |
| Required fields | Link, status code, review result, replacement link |
Collection and judgment
- Export the link list. Export all evidence links from the source register, de-duplicate, and keep the source page.
- Send batch requests. Record status codes, redirect destinations, and response times, and mark timeouts separately.
- Review by hand. Open each anomalous entry and separate real failures from temporary faults.
- Handle the result. For real failures, add a replacement link or delete the claim; for redirects, update to the destination address.
- Archive the record. Keep the inspection list so repair progress can be compared.
Data sources and record fields
Data comes from inspection records, with fields including link, status code, redirect destination, review conclusion, action taken, owner, and date.
Actions must be tied to the content ledger; when a claim is deleted, the page source block must be updated at the same time.
Limitations and misreadings
- Batch requests may produce false positives due to rate limits or protection policies, so manual review is required before counting.
- Pages that require login or interaction cannot be judged by automated detection.
- A working link does not mean the content is unchanged; semantic drift requires human reading.
- A low broken rate does not mean evidence is sufficient; when there are too few entries, the ratio has no reference value.
Relationship to other metrics
The broken rate complements primary source coverage: the former looks at whether sources still exist, the latter at the structural level. Combined with evidence freshness, it separates failure from expiration.
Related sections: metric index, method index, glossary index, research index, checklist index.