Synapse builds downward: fewer frameworks, less cloud overhead, tighter deterministic control. Today that means a self-hosted runtime for evaluators, reward functions, policy gates, and other bounded AI-generated logic. A supported Python subset compiles to WebAssembly, runs in a tiny sandbox, and can be verified with Z3. On top of that runtime, the governed control-plane shell adds redaction, receipts, replayable proof packs, and bounded escalation for design-partner workflows. The hosted endpoint is a preview. The product is the stack you deploy yourself.
Self-hosted deterministic execution for bounded generated logic. Start with evaluators, reward functions, policy gates, and auditable transforms.
Fewer layers. Lower overhead. Focused execution surface.
The same stack also supports the deeper R&D vector: transpilation, formal verification, receipts, model-improvement loops, energy efficiency, and the long-horizon sovereignty path toward bare metal.
Downward means more control, more assurance, and eventually more sovereignty.
Use the hosted preview to evaluate the runtime before self-hosting. Production deployments run on your own gateway. This demo shows the same bounded execution model on a public evaluation surface.
Benchmarks and research claims are tied to documented hardware, journaled experiments, and reproducible commands. Start with the runtime, then go deeper into the research.