Atlas 0.2.29 puts a named approver inside the agent’s answer

Atlas 0.2.29 changes approval from a workspace-level event into claim-level provenance. The release adds a published_by field, a scoped approval operation and approver details in searchAtlas results. That matters because Atlas’s product contract says an attested fact should carry its source, date and the name of the person who approved it—not merely the identity of the person who originally said it. The merged implementation describes that earlier gap explicitly.

Approval is now an explicit, bounded operation

The new POST /api/v1/admin/brain-facts/approve endpoint accepts an explicit list of one to 200 fact IDs. It runs the existing promotion adapter and transaction rather than creating a second path that can publish records. An empty list returns an error instead of widening into an unscoped publish, and a request that cannot be attributed to a person is refused. The same implementation says already-published, retracted, missing or cross-workspace IDs do not silently become eligible; clients must compare requested IDs with the returned promotedIds list.

That is a useful API design for analytics agents because “a human reviewed this workspace” is weaker evidence than “this person approved this particular claim.” A result consumer can now distinguish the recorded author—the person whose words supplied the claim—from the approver who accepted it for serving.

The identity travels with retrieval

Atlas now returns an approval object with search results. Approved facts can include the approval time and an approver resolved from the live user record; local deployments can identify a local-operator. If a user lookup fails, the documented behavior preserves the approval state with a null display name rather than failing the entire search. Atlas’s README separately describes its three evidence classes: live “Surveyed” data, human-approved “Attested” facts and raw “On the record” material.

The implementation also repairs a related date hole: one promotion branch could set a fact to published without stamping published_at. The change makes both promotion branches write the approval time and approver together, with tests covering that shared property.

There are limits. The pull request says the release does not add a dedicated web interface for the new approval verb, does not backfill older rows and does not yet include a route-level test for POST /approve; lower-level adapter and actor-mapping tests cover the underlying logic. The v0.2.29 release shipped the change alongside MCP demo setup and documentation updates.

The practical takeaway is narrow but important: provenance should identify both the source of a claim and the authority that allowed an agent to repeat it. Atlas now makes that distinction part of the retrieved answer rather than leaving it implicit in an audit trail.