Sometimes the asset already has what the check needs, for example the row count it just wrote to the warehouse, and loading it again would be wasteful. Declare the check on the asset with check_specs=[dg.AssetCheckSpec("no_unexpected_negatives", asset="clean_transactions")] and return the result with the materialization: dg.MaterializeResult(metadata={...}, check_results=[dg.AssetCheckResult(passed=...)]).
The check appears on the asset page like any other, and blocking=True on the spec works the same way. Every declared spec must get a result whenever the asset runs; if one is missing, the step fails with an error naming the check output that was not returned.
Prefer a separate @dg.asset_check when the check reads the stored data independently (so it can be run alone later), and check_specs when the check is a by-product of writing the data.
The check result travels with the materialization.