Quality & methodology
How we measure, and what we won’t claim.
Bearing data is only useful if you can tell how good it is. This page defines every measure we will report, and says plainly which ones have not been measured yet.
Current state (9 October 2026): no coverage, completeness or accuracy figures are published, because no verified, publishable MyBearings dataset has been evaluated yet. When figures appear here, each will carry its definition, denominator, sample and date.
Measures
Seven measures, never blended
Kept separate on purpose. A value can be complete but unsourced, sourced but mapped to the wrong field, or correct for the wrong product.
| Measure | Definition | Denominator | Status |
|---|---|---|---|
| Product coverage | Unique manufacturer products, by manufacturer and family | Counted separately from supplier listings | not yet measured |
| Applicable-field completeness | Populated ÷ applicable attributes | Excludes not-applicable; unknown applicability reported separately | not yet measured |
| Source support | Share of populated values with ≥1 recorded source assertion | Populated values | not yet measured |
| Accuracy | Agreement with expert review on a held-out sample | Sample size and scope stated | not yet measured |
| Identity precision | Accepted matches that are correct; false-merge rate | Accepted matches (candidates excluded) | not yet measured |
| Open conflicts | Fields with disagreeing supported values awaiting review | Applicable fields | not yet measured |
| Freshness | Time since last successful source verification | Per source and attribute | not yet measured |
Principles
Rules we hold ourselves to
- 01
Manufacturers are primary evidence
Manufacturer documentation is the primary authority for its own products. We reconcile sources; we do not claim to replace them.
- 02
Source hierarchy per attribute
One appropriate source can suffice. We cross-check more where there is conflict, ambiguity or high impact — not by blanket rule.
- 03
Conflicts are preserved
Disagreeing supported values stay as separate assertions; review selects one with recorded reasoning and history.
- 04
Scope is explicit
Product, supplier listing and physical lot facts are different. Lot origin never overwrites a product field.
- 05
Confidence is not probability
Any confidence value is a method-specific estimate. Calibration against expert review is separate work and will be published separately.
- 06
No evaluation on training data
Accuracy is measured on held-out records that were not used to develop the rules being evaluated.
Corrections
How corrections work
Errors will happen. What matters is that they are easy to report and their fixes are visible.
Report
Send the record, the field and your evidence through the contact form.
Check
We review the sources behind the current value and yours.
Record
Accepted corrections become new assertions; the old value stays in history.
Publish
Corrections are reflected in the next data version with a change note.
Read more
Method in practice
Bring one messy category.
Leave with a clear next step.
Share what your catalog looks like today. We will tell you what a scoped assessment would cover, what it would not, and what we would need from you.