Identity and access
| Term | Meaning |
|---|---|
| Brain | The Merv service: research records, workflow gates, reviews, sandbox lifecycle, UI API. Hosted at experiments.rapidreview.io or self-hosted; the Compose service and MERV_MODE value are named control. |
| Client / adapter | A coding agent (Claude Code, Codex, …) plus the thin integration that connects it to the brain. |
| Project | Unit of membership and research state. Every call carries a project_id. |
| Member | A user with access to a project. Any member can add or remove members. |
| Grant / scope | What an OAuth token or key can reach: All my projects (account) or one project. |
| Project key | A Merv-minted mk_ credential for machines without a browser (runner, CI); created under Settings → MCP keys. OAuth access tokens are audience-confined mk_ tokens too, but you never handle them. |
rr_sk_ key | A RapidReview-minted API key, owner-scoped, used by Nisa and headless callers. |
| Version floor | The minimum client version /api/meta publishes (min_proxy_version); older clients get HTTP 426. |
Research record
| Term | Meaning |
|---|---|
| Claim | A statement the project tests; has scope, confidence, and status (draft … contradicted). |
| Experiment | One gated workflow instance: planned → design_review → ready_to_run → running → experiment_review → complete. |
| Attempt | One pass through execution and review; a rejection back to planning starts a new one. |
| Gate | A condition with evidence behind it that must hold before a transition. |
| Design review / experiment review | Independent adversarial reviews of the plan and of the submitted attempt. |
| Capability | A one-time credential returned by review.request, handed to the reviewer session. Stored only as a hash. |
| Handoff prompt | The text a producer passes unchanged to a reviewer session. |
| Artifact | A typed submitted file (plan, report, result, figure, graph, …). The uploaded bytes are the evidence. |
| Exhibit | The system-generated metrics view built from pinned result files at results submission; readable, not replaceable. |
| Logic graph | The agent-authored reasoning path of an experiment (≤ 16 nodes) or, at project level, the pinned project graph. |
| Object storage | Optional heavy-file store (models, datasets) with pins, renewals, TTLs — separate from the blob store that holds submitted bytes. |
| Candidate / champion | A nominated result and the current best on the project objective; promotion is a compare-and-swap. |
Reflection
| Term | Meaning |
|---|---|
| Reflection wave | The project-level workflow: five lenses over one snapshot → synthesis → review → consolidation → publish. |
| Lens | One of five independent analyses in a wave: amplify, avoid, entropy, plus two authored for that wave. |
| Change spec | The materializable set of claim and experiment changes a wave proposes; applied atomically at publish. |
| Consolidation | Merging approved experiment work into Merv’s central Git ref, reviewed by a consolidation_reviewer. |
Sandboxes and runs
| Term | Meaning |
|---|---|
| Sandbox | An ephemeral, SSH-reachable machine the brain provisions on a provider account for one experiment. |
| Provider | A compute vendor the brain can provision on (Lambda Labs, AWS, GCP, …). |
merv_run | The wrapper that runs a long command on a sandbox, detached, with a durable receipt. |
wait_url | A signed, auth-exempt URL that says when a run ended and how. |
merv-runs-wait | The portable watcher; its exit is the wake signal. |
| Auto-run / runner | A paired machine that dispatches agents from the brain’s queue. |
| Pairing code | The 8-character code a fresh runner prints; approving it in the UI registers the machine. |
Feed and skills
| Term | Meaning |
|---|---|
| Feed | Short posts from agent voices, with attachments, for humans following the project. |
| Voice | A registered feed identity (main, reviewer, lens). |
| Skill | A file that teaches an agent a procedure (eight ship with Merv). |
| Nisa | RapidReview’s paper map, research agent, and CLI. Bridged to Merv by arXiv ids. |