Research

retflo is a typed argument graph: 66 argument nodes across seven domains, connected by six edge types (flow, resolution, premise, escalation, retreat, redirect). Each node carries the case against it, in its strongest form. The repo is public and the whole structure is machine-readable. Everything on this page is downloadable today.

Computational argumentation

The graph is a formal object you can compute over. Nodes, edges, and objections are all in JSON, served whole through the API and carried as plain files in the repo. Edge types are part of the data, so structural questions are queries: path lengths, cycles, edge-type distributions, which objections attach to which claims.

A conversation can be projected onto it: a real debate becomes a sequence of named nodes, and recurring patterns become measurable. We do not publish a conversation corpus. The graph is the structure to project one onto. The node browser shows what the units of that mapping look like.

An evaluation target

The graph is a fixed reference a model's output can be checked against. Given a political claim, does a model route it to the node that answers it? Given a node, does it reproduce the objections at the strength recorded there? Checks like these need a stable answer key, and the graph is one: public, typed, and versioned in git.

The interfaces you would test are the ones anyone uses: the agent skill, the local MCP server, and the Markdown the site serves to LLMs. The AI page covers them. Nothing in the setup requires our involvement.

Other domains

The format is separable from the subject. Typed edges, the case against each argument stored alongside it, claims sorted into fact, frame, or contested: none of that is specific to politics. Any contested field where the same arguments recur could be mapped the same way. The repo is open under the RCCL, and a fork for another domain is a use we would like to see.

Finding gaps

The graph is also a target. A well-formed objection that does not route to an existing node is a gap, and finding one is a contribution. When the framework loses on the merits, the correction lands in the public record. The visualizer shows the whole structure if you want to look for thin spots.

Getting started

The graph ships in four forms: HTML on this site, Markdown for LLM context, JSON through the API, and the interactive visualizer. Terms are in the RCCL: free for individuals, co-ops, and democratic organizations, licensed for commercial use. Start with the node index.