Problem
A useful correction can get buried in a long conversation. Replaying the whole conversation for a new task can also bring in another team’s context or give an unreviewed suggestion too much authority.
Technical decision
I separated observation, review and retrieval. New material retains its origin as a candidate. A separate review step controls promotion, and retrieval is scoped by tenant, role, task and a context budget.
What I built
I built SDK and API workflows for submitting evidence, compiling candidates and requesting context. Graph navigation and curation views let a reviewer inspect relationships and decisions while working in the selected project.
Evidence
A documented local integration test used synthetic evidence with the real producer, bridge and knowledge service. It checked delivery across that boundary. Product walkthroughs separately showed graph filtering and navigation into curation. Downstream effects on task quality and cost remain unmeasured.
