Follow-up to #367 (part of #208). Edge trust propagation shipped (extracted edges carry their source chunk's trust, surfaced on recall; client-prompt hardening guidance added to the security guide). Remaining:
- Node-level trust — a node extracted purely from untrusted content could also carry a trust marker. It's murkier than edges because a node (entity identity) can accrue from many sources of differing trust: a resolution is needed (e.g. a node is "untrusted" only if ALL its live edges are untrusted, or track a per-node provenance set). Design + implement + surface on recall/get_node.
- Per-call add_document trust — the MCP
add_document tool currently ingests as untrusted (the fail-closed default). Add an optional trust arg so a client that vouches for a specific page it pushes can mark it trusted per call, without a per-source config entry.
Both unit/DB-testable without live infra.
Follow-up to #367 (part of #208). Edge trust propagation shipped (extracted edges carry their source chunk's trust, surfaced on recall; client-prompt hardening guidance added to the security guide). Remaining:
add_documenttool currently ingests as untrusted (the fail-closed default). Add an optionaltrustarg so a client that vouches for a specific page it pushes can mark it trusted per call, without a per-source config entry.Both unit/DB-testable without live infra.