PulseGraph currently supports one deployment model: a single user running the frontend and backend on the same machine.

Training resources and attached model source are executable Python. Resource preview, training, source attachment, and forward replay may import modules and run their top-level code. Only load code you wrote or reviewed.

The current safeguards are defense in depth for this local workflow:

• the provided server commands bind to 127.0.0.1;
• the API rejects non-loopback clients and non-local browser origins;
• Python execution routes require the explicit X-PulseGraph-Trust: trusted-local-code marker;
• weights-only artifact inspection uses restricted PyTorch loading;
• live-run memory, subscriber queues, and event persistence have bounded or locked write paths.

The trust marker is not authentication, and a local subprocess is not a security sandbox. PulseGraph does not currently isolate uploaded Python from the host filesystem, network, environment variables, or the permissions of the user running the backend.

Do not expose the current backend to a LAN, public network, reverse proxy, shared workstation, or multi-user environment. CORS is not an authorization boundary, and runs do not have per-user ownership.

Before a shared deployment is supported, PulseGraph needs authenticated users, run ownership checks, isolated worker processes or virtual machines, resource quotas, and a durable transactional event store.