Docs menu
Quality and validations
The trust score, its six dimensions and the rules that guard your data.
In short
Each analysis gives a dataset a quality score from 0 to 100, built from six dimensions. The score shows which dimension pulls it down, and a severe problem caps it so it cannot be averaged away. Validation rules turn your business expectations into checks that run at every analysis.
The quality score
The score tells you how far you can rely on a dataset today. It appears on the Quality tab and in the dataset header.
| Score | Level |
|---|---|
| 80 to 100 | Excellent |
| 60 to 79 | Good |
| Below 60 | Needs improvement |
Next to the score you find:
- the change since the previous analysis;
- how many dimensions it rests on, for example "Based on 5 of 6 dimensions";
- the reason, when the score is capped.
The score is computed at each analysis. An analysis runs:
- when someone clicks Run an analysis on the Quality tab;
- after a transformation builds or rebuilds the dataset;
- when a data pipeline requests it through the API.
Each analysis also runs the validation rules and checks the dataset's alerts.
The six dimensions
| Dimension | Weight | What it measures |
|---|---|---|
| Validity | 30 | Your validation rules, weighted by severity |
| Completeness | 20 | Filled values per column, compared with each column's own history. Placeholders such as "N/A" or -1 count as missing. |
| Freshness | 15 | Age of the newest data against the expected update cadence |
| Volume | 15 | Row count change against the dataset's usual changes. A table that only grows is not penalised for growing. |
| Uniqueness | 10 | Duplicated rows and duplicated keys |
| Stability | 10 | Column drift, schema changes and the share of anomalous rows, against the dataset's history |
A dimension without signal is left out instead of being scored 0. For example, validity has no signal until the dataset has a validation rule. The score is then computed on the remaining dimensions.
Expected freshness comes from, in this order:
- a freshness validation rule, if you set one;
- the update cadence learned from the dataset's history;
- otherwise, 48 hours.
Caps
Some problems are too serious to be averaged with good results. They cap the score:
| Problem | Maximum score |
|---|---|
| A critical validation rule fails | 59 |
| The primary key has duplicates | 59 |
| The table is empty | 40 |
| The schema changed in a breaking way | 40 |
A capped score always stays below Good or Excellent, whatever the other dimensions say.
Validation rules
A validation rule states what the data must respect. Rules are grouped in suites, on the Validations tab. Examples:
- the email column must never be empty;
- revenue must be positive;
- the status must be one of an approved list of values;
- the table must hold between 1,000 and 50,000 rows;
- the newest row must be less than 24 hours old.
Rules check columns (values, lengths, ranges, patterns), statistics (mean, median, percentiles), the whole table (row count, freshness, schema changes, duplicated rows), or a custom SQL query.
Each rule has a severity, which sets its weight in the validity dimension:
| Severity | Weight | Use it for |
|---|---|---|
| Critical | 5 | A rule whose failure makes the data wrong for decisions. A failure caps the score at 59. |
| Warning | 2 | A problem worth fixing, which does not invalidate the data |
| Info | 0.5 | A signal to follow |
The owner, an editor or a platform administrator can create and edit rules.
AI recommendations
On silver and gold datasets, the AI can suggest validation suites from the dataset's profile. You review each suggestion, then accept it to add it to the dataset's validations, or dismiss it. Suggestions need at least 70% of the dataset and its columns to be documented.
Early warnings
Each analysis also detects problems by itself, each with a severity:
- schema changes (removed, retyped or added columns) and empty tables;
- volume drops and stale data;
- too many empty values, and placeholder values such as "N/A", blank strings, -1 or 9999;
- duplicated keys and duplicated rows;
- dates in the future, columns full of zeros, constant columns;
- a change in a column's distribution since the previous analysis;
- an unusual share of anomalous rows, with the columns that explain them.
These warnings need no setup. To be notified, create an alert.
AI summary
On silver and gold datasets, an AI summary explains each analysis in plain language: what changed since the previous analysis, the main issues and their impact, what the anomalous rows have in common, and what to do next. A suggested validation rule can be added in one click.
- The summary is built from aggregated statistics only, never from raw rows.
- It is only generated again when the analysis actually changed.
FAQ
My dataset has no validation rule. Is its score reliable? It rests on fewer dimensions: validity is left out. Add a few rules to make the score more meaningful.
Why did the score drop although nothing failed? Open the score breakdown: freshness often drops when an update is late, and volume when the row count moves more than usual.
Why is the score capped at 40? The table is empty, or its schema changed in a breaking way. The reason is shown next to the score.