WorldGraph adds bounded capture playback to its digital twin

What WorldGraph is

WorldGraph is two things in one repository. The first is a Rust library that models a building as a typed graph: rooms, zones, sensors, people and beliefs about what is happening, each with provenance attached and the ability to forecast occupancy.

The second is an AI coding agent, published on npm as worldgraphs, that walks you through architect, implement, review and test while you build a twin of your own space.

What changed

Commit e72a1d5 added bounded Gaussian capture playback with native provenance and validation, so a recorded capture can be replayed through the graph with its origin intact instead of being treated as anonymous data.

Follow up commit 9b1c79c clarified the verified artifact publication path, and an earlier commit in the same series made the lab export the Rust world state synchronously.

Get started

Node 18 or newer for the agent side. Nothing to install first:

npx worldgraphs init
npx worldgraphs doctor

Expected result: init wires the agents into your AI host, and doctor prints a health check of what is present and what is missing.

For the Rust library:

cargo add wifi-densepose-worldgraph

The read only MCP tools run from a source checkout with node bin/cli.js mcp start. They create mission and validation plans; they do not write to your world state.

Use it today

Practical case: an office that wants occupancy forecasting without cameras. Input is sensor feeds and a floor layout. Workflow is model the space as zones, attach sensors, replay a capture, check the beliefs the graph produces. Output is a provenance tracked occupancy estimate.

Acceptance test: replay the same bounded capture twice and confirm the resulting world state matches. Playback that is not deterministic is not a usable baseline.

Push it further

Experimental commentary. Pair the twin with RuView sensing and you get a building that maintains its own model rather than one a consultant draws once and nobody updates.

Limitation: forecasting quality depends entirely on sensor coverage, and nothing here validates a sparse deployment. Falsifiable test: remove a third of the sensors from the graph and measure how far the occupancy forecast degrades.