Skip to main content

Entity Graph

The entity graph is the interactive map of who's who in a simulation and how they related to one another. It's one of the four result views, rendered as a node-and-edge diagram you can pan, zoom, and click into.

Nodes: the agents​

Each node is one agent (one entity the simulation ran on). Beyond its name and type, a node carries activity stats computed from the chat log — how many actions it took, how many posts it authored, and how many rounds it was active. This lets the graph show at a glance who drove the conversation and who stayed quiet, not just who was present.

Edges: two kinds of relationship​

Edges come from two complementary sources, and both appear on the same graph:

  1. Interaction edges are derived deterministically from the chat log. When one agent follows, replies to, votes on, reposts, or mentions another, that becomes a directed edge labeled by the interaction type (a negative vote becomes a dislike edge, a positive one a like). Repeated interactions between the same pair collapse into a single edge whose weight is the count, so a thick edge means a strong, repeated tie. Mentions are detected both from @handle references and from full entity names appearing in post text, and post-targeted actions are resolved back to the post's author so a reply links the two agents, not the agent and a post ID.

  2. Semantic relationship edges are derived by an LLM that re-reads a sample of the transcript and proposes typed edges between entities — relationships like supports, disagrees with, or responds to, each optionally carrying a short fact describing the relationship. These typed edges are merged alongside the interaction edges. The extraction logs its raw output and retries once on a parse failure, and any relationship whose endpoints don't map to known entities is dropped rather than guessed.

How it's presented​

The graph ships as a snapshot of nodes, edges, and metadata (totals, the number of rounds, and the set of entity types) that the frontend renders with Cytoscape.js. Together the two edge types give you both the mechanical structure of the run (who interacted with whom, how often) and its semantic structure (what those interactions meant) on a single canvas.