Proposal: governed collective-decision reference application for LangGraph

Problem

LangGraph provides graph execution, state, routing, persistence, and human-in-the-loop primitives. A multi-agent decision graph can still accept malformed carriers, merge incompatible candidates, allow a reviewer to create a new answer, or disclose more state than a role needs.

Proposed reference application

Rulora’s collective-decision example demonstrates an application-level control chain:

independent model outputs
  -> deterministic carrier Recovery
  -> semantics-preserving Adapter
  -> Program-frozen candidate IDs and hash
  -> Reviewer selects IDs only
  -> Program expands and validates
  -> auditable JSON delivery

LangGraph handles graph execution. Rulora adds application-level decision contracts and role permissions; it does not replace LangGraph’s checkpointer, routing, retry, or interrupt model. Program allowlists limit the information disclosed for independent decisions, bounded broadcast revision, and constrained review.

Reproduction

The example now lives in the Rulora main repository. With Node.js 20 or newer:

git clone https://github.com/Buffalo2024/Rulora.git
cd Rulora/examples/collective-decision
npm install
npm run example:langgraph
npm run test:minimal
npm run eval:reliability

No model key is required for these commands. The fixtures cover valid JSON, fenced JSON, a missing-comma carrier, ambiguous or invalid values, and invented Reviewer IDs. The example includes a langgraph.json graph configuration.

The checked-in corpus contains 10 carrier/semantic fixtures, 3 constrained-Reviewer fixtures, and 3 model-role disclosure views. On those fixtures only:

  • accepted business-value preservation: 1.0
  • ambiguous or invalid output rejection: 1.0
  • forbidden Reviewer selection rejection: 1.0
  • model-role view leakage: 0.0

These are regression results for a small, inspectable corpus, not claims of model accuracy or universal reliability.

Questions for the LangGraph community

  1. Is this most useful as a reference application, template, guide, or community integration?
  2. Should the reusable boundary expose graph nodes, node wrappers, or a graph factory?
  3. Which LangGraph-native persistence and interrupt example would make the application most useful without duplicating framework responsibilities?

Limits

The sample uses synthetic decision data. It is not a validated decision system for real-world business use. Feedback supported by failure fixtures, measurements, or minimal reproductions is especially welcome.