graph-reviewer-nodelisted

# Graph Engineering: The Reviewer Node **Source**: AI Builder Club's graph-engineering guides name "give the reviewer node authority — separate, read-only verifier" as a specific implementation checklist item. **Academic corroboration**: "Verified Multi-Agent Orchestration" (arXiv 2603.11445) implements this exact idea as "orchestration-level verification" — an independent model assessing whether the *collective* result of a multi-agent run adequately addresses the original goal, decoupled from any individual agent's self-assessment — and measures a real quality gain from it (completeness 3.1→4.2, source quality 2.6→4.1 on a 1-5 scale, versus a single-agent baseline, on 25 market research queries). ## Why this is the highest-leverage single node in most graphs The Researcher → Writer → Reviewer starter pattern (see `graph-node-and-edge-design`) only produces a real quality gate if the Reviewer node genuinely has the two properties its name implies: 1. **Read-only** — it doesn't also produce the work it's judging. A node that both writes and reviews its own writing has the same self-verification problem `loop-verifier-design` names for single-agent loops, just relocated into a graph. 2. **Authority** — its verdict actually controls the routing edge. A "reviewer" node that produces commentary nobody's conditional edge checks is decoration, not a gate. ## What real authority looks like, concretely The conditional edge from Reviewer needs to branch on the Reviewer's actual