Grounding & Evidence

Surfaced as 'Sources', built as an engine

Internally this is the Brain's organizational intelligence engine; in the dashboard it's just the Brain's "Sources" tab — the engineering name is deliberately kept out of the UI. Its job is to let organizational claims carry inspectable evidence rather than floating as unsourced assertions, and to give those claims a real lifecycle instead of treating memory as flat and equally trustworthy.

Claims have a workflow, not just a value

A claim isn't written and forgotten — it moves through actual states. GET /v1/grounding/claims/:id returns a claim with its evidence enriched by citation summaries (title, a native link back to the source, and whether that evidence version is still live — i.e., the source hasn't since changed underneath it). Claims can be explicitly approved or rejected (POST /claims/:id/approve / /reject), and when two claims conflict, GET /claims/conflicts surfaces the disagreement for explicit resolution rather than one silently overwriting the other.

Evidence is versioned

The evidence ledger tracks versions of a source over time, not just its current state — so a claim can be traced to precisely the version of a document or page that supported it, and the system can tell you whether that specific version is still the live one or has since been superseded.

Grants scope what an agent can see

GET /v1/grounding/agents/:agentId/grant reflects that grounded organizational knowledge isn't automatically visible to every agent — access is granted per agent, so a specialist only sees the grounded context it's actually been given permission to draw on.

Continue