Skip to main content

Research Alpha architecture

EXEPERT Research Alpha v0.4.0-alpha.1 is an unreleased, local-first proof that one deterministic game environment can move between visual play, headless execution, Python training, and replay without changing simulation semantics. Researchers bring the broader training stack; EXEPERT supplies the environment contract, execution adapters, run artifacts, and inspection surface.

The DOM-independent kernel is authoritative. Phaser renders snapshots and may interpolate presentation, but does not own collision, reward, spawn, time, or terminal state. The Node worker imports the same kernel and reserves stdout for versioned protocol responses. Python communicates only through that protocol.

Trust and privacy boundary​

  • Training is local and every generated artifact stays under ignored .exepert/runs/<run-id>/ unless the user explicitly copies it elsewhere.
  • The browser application remains usable when the loopback service is absent.
  • The service binds only to 127.0.0.1, accepts known local origins, requires a per-process session token, validates input, and launches one predefined training command. It cannot execute arbitrary shell input.
  • Artifact access is confined to the run root. Repository secrets, .env, and unsupported file types are not exposed.
  • Worker stderr and service errors are sanitized. No hosted Supabase, Vercel, telemetry, upload, billing, or external model service is required.

This alpha promises repeatable environment transitions for the same environment version, seed, and actions. It does not promise bit-identical PPO weights across operating systems, hardware, or numerical-library builds.