Merv
Nisaproject_id on every call. project(action="overview") returns the whole picture: claims, experiments (including terminal ones you shouldn’t recreate), the latest published reflection, the literature summary; project(action="create") makes one (not available to external keys).
Claims
A claim is a statement the project is trying to establish, refute, or refine. Each carries ascope, a confidence (low · medium · high), and a status:
Agents create claims with
claim.create and move them with claim.update; reflection waves apply approved claim changes atomically at publish.
Two graphs
Experiment logic graph (graph.json, submitted with role graph): the reasoning path of one experiment — questions, decisions, pivots, consequences, lessons — as a DAG of at most 16 nodes. Agent-authored, not generated: choosing what mattered is the authorship. Not a pipeline diagram, not an event log.
Project graph: produced by a reflection wave’s synthesis, reviewed, and pinned at publish. It’s the project-level logic — how claims and experiments relate and where the next phase goes. The UI renders both.
Champion and candidates
Projects with an objective track a champion. The manager agent nominates promising output withcandidate.submit (an immutable artifact or storage object, or an experiment_workspace for the evaluator to stage); candidate.list shows candidates, the append-only promotion history, and the champion; candidate.promote is a compare-and-swap with a required reason.
Literature
The project keeps a living literature review (litreview.view / litreview.edit, one targeted change per call) and a papers ledger (litreview.cite, deduplicated across arXiv/DOI/URL forms and linked to sections, experiments, or claims). Plans cite papers in portable form — arXiv:2401.12345, a DOI, or a URL. See Using Nisa with Merv.