# Is the data actually right?

Prompt · Workflows · DBR01

Prove the numbers before fixing anything.

## Prompt

```text
Is the data actually right? Prove the numbers before fixing anything.
Work through this in order and check in with me after each step. Stay read-only until I approve a change.

1. Start: PG01 when a figure disagrees with its source; PG02 when everything validates but still looks wrong.

Can we independently prove these numbers? (PG01)
Take a representative set of important figures from this database and independently reconstruct them from their original sources. Don't rely on the existing calculation code to verify itself. Find discrepancies in definitions, source selection, units, dates, classifications, and relationships. Determine which errors are isolated and which come from shared logic, then rank corrections by analytical impact.

Find data that is internally consistent but wrong (PG02)
Look for data that passes the database's constraints and ordinary validation but is still analytically wrong. Test cross-table relationships, accounting identities, historical continuity, plausible ranges, source coverage, and independent calculations. Focus on contradictions that existing checks would miss, and distinguish genuine errors from legitimate differences in definitions or reporting periods.

2. If needed: PG03 for identity gaps; PG04 when the meaning of past records matters.

Find where records silently change identity (PG03)
Trace how real-world entities are identified, matched and connected across imports, source documents and database tables. Find cases where distinct entities are merged, identical entities are duplicated, or historical relationships become disconnected. Identify the matching rule responsible and the affected population before proposing corrections.

Reproduce what the database believed in the past (PG04)
Choose important historical records and determine whether the system can reproduce what was known and reported at a specific past date, using the definitions and evidence available then. Distinguish later source corrections from later classification changes. Find where current logic silently rewrites historical meaning and recommend the smallest reliable solution.

3. Change: Fix only the confirmed shared causes, and add SQL or metric tests for them.

4. Done when: Recomputed examples that now match, an estimate of how much data was affected, and any source limits stated plainly.
```

## Related

- [Can we independently prove these numbers?](https://www.atlasscope.dev/prompts/can-we-independently-prove-these-numbers): Important figures may be wrong even though their lineage is documented.
- [Find data that is internally consistent but wrong](https://www.atlasscope.dev/prompts/find-data-that-is-internally-consistent-but-wrong): Every check passes, but totals or outputs still look implausible.
- [Find where records silently change identity](https://www.atlasscope.dev/prompts/find-where-records-silently-change-identity): Sources use aliases, changing identifiers, share classes or inconsistent names for the same thing.
- [Reproduce what the database believed in the past](https://www.atlasscope.dev/prompts/reproduce-what-the-database-believed-in-the-past): Past reports, backfills, restatements or changed classification rules matter.

Source: https://www.atlasscope.dev/prompts/is-the-data-actually-right