# Grading method (draft)

Data Quality (DQ) ratings come from reproducible tests against the source. No ratings are issued yet.

## Six dimensions

1. **Provenance** (weight 20%): Is the publisher the original holder, and can every record be traced to it?
2. **Freshness** (weight 15%): How long after a change at the source does it appear in the API?
3. **Accuracy & validity** (weight 25%): Do sampled records match the source field by field, in the documented formats and schema?
4. **Completeness** (weight 15%): What share of expected records and fields are present?
5. **Licence clarity** (weight 10%): Can an agent tell, from the response alone, what it's allowed to do with the data?
6. **Reliability** (weight 15%): Uptime, latency, error rates and unannounced schema changes, measured from outside over time.

## Scale

Each dimension is scored 0–100. The DQ score is their weighted average, shown as a band from A to E, for example "DQ A · 88".

- **DQ A** (85–100): Excellent. Dependable for unattended agent use.
- **DQ B** (70–84): Good. Minor, documented gaps.
- **DQ C** (55–69): Fair. Usable with checks on the weaker dimensions.
- **DQ D** (40–54): Weak. Significant gaps found in testing.
- **DQ E** (0–39): Poor. Not fit for unattended agent use.

- A Licence clarity or Provenance score below 50 caps the rating at C, however strong the rest is.
- A rating expires 90 days after its last test and shows as Ungraded until it's retested.

Reports from users queue a test. Only tests move a rating.