# Infer and confirm

What Datapace reads, what it infers with a confidence score, inferred versus validated, and how a few working sessions turn proposals into a confirmed model.

## What Datapace reads

Datapace is designed to read within the scope agreed with your team, and to read three kinds of signal:

- **Schema metadata.** Tables, columns, types, constraints, keys, views, and the comments people left behind.
- **Workload signals.** Which queries run, how often, how long, on which tables, joined how.
- **Sample structures.** The shape of values (formats, ranges, distinct counts), which is how an email column or a status code announces itself.

Workload signals are what a schema alone cannot give: a join that runs ten thousand times a day is a relationship whether or not a foreign key declares it, and a table nobody has read in a year says something no comment does.

## What it infers

From those signals Datapace proposes the meaning facts of [the graph](/docs/concepts/the-graph): entities, measures, relationships, semantics, and lineage. Every proposal carries a **confidence score**, a number between 0 and 1 that says how much of the evidence points the same way.

| Signal | What it suggests | Typical confidence |
| --- | --- | --- |
| A column named `email` whose values all match an address format | Personal data, email | High |
| Two tables joined on the same pair of columns in most of the workload | A relationship | High |
| A free-text `notes` column with occasional phone numbers | Personal data, to review | Low |
| A table named `tmp_customers_2019` with no reads in a year | A stale copy | Medium |

The table is illustrative. The point is that the score is on the record and that a low score is a question for a person, not a guess that slips through.

## Inferred versus validated

Every meaning fact is in one of two states:

- **Inferred.** Proposed by Datapace, with its score. Visible, usable for review, not yet trusted.
- **Validated.** Accepted or corrected by one of your experts. Part of the confirmed graph.

Agents draft their work on the confirmed graph. AI systems reading through [governed context](/docs/concepts/governed-context) see validated facts, and see inferred ones only as such. Nothing enters the confirmed graph unvalidated.

## The working sessions

Confirmation is a few sessions with the people who know the data, not a project. Datapace sorts the proposals so that a session covers what matters first: the entities most of the workload touches, the columns flagged as personal data, the relationships with the most joins behind them, the measures dashboards already report.

In a session an expert accepts a proposal, corrects it, or marks it for someone else. Each decision is recorded with who made it and when. See [documenting an estate](/docs/guides/documenting-an-estate) for how a first pass runs.

## Living documentation

Once confirmed, the graph is the documentation, and it stays current because the signals keep arriving. A new table appears as an inferred entity waiting for review; a query pattern that changes shows up as a relationship to confirm; a column that starts carrying addresses is flagged again. The Documentation agent proposes the updates and your team approves them, the same way it approves everything else.