Isolation Model

Synapse executes untrusted code in a deterministic, sandboxed WebAssembly arena. No filesystem. No network from guest code. Memory wiped between every request.

EXECUTION SANDBOX

Each execution runs in a 64KB Wasm linear memory arena.

No filesystem accessguest code cannot read or write files
No network accessguest code cannot open sockets or make HTTP requests
No syscallsWasm modules interact only through explicit FFI host functions. Most are reserved stubs in the public runtime — only print is active.
Memory isolationeach module has its own linear memory, invisible to others
Arena resetmemory is wiped to zero between every request
Fuel meteringinfinite loops are terminated after a configurable instruction budget

HOST / GUEST BOUNDARY

Guest Wasm Module
64KB linear memory
No fs, no net, no syscalls
13 FFI Functions (Host-Side Only)
Active: print
Reserved: tcp, crypto, spawn, kv, sql
Reserved functions are defined in the FFI spec but return stub values in the public runtime. They are not exposed to guest code on the demo gateway.
Host: Rust Gateway
wasmtime 43, fuel metering
moka cache, arena allocator

DEPLOYMENT MODEL

The gateway is a single ~15 MB Rust binary with zero external runtime dependencies.

No GPU requiredruns on commodity CPUs
Deployable behind firewallsair-gapped networks supported
No telemetryno phone-home, no external calls
Reduced runtime supply-chain surfacesingle statically-linked binary
Long-term stabilityif it runs today, it runs identically in ten years

OPERATIONAL SAFEGUARDS

Hosted demo throttlingthe public demo surface is IP-throttled and intended for evaluation, not shared production traffic
Self-hosted edge controloperators are expected to terminate auth and TLS at their own edge while the core gateway stays focused on deterministic execution
Fuel meteringprevents resource exhaustion from runaway programs
Thread-per-core modelno contention between requests
No data persistencestateless by default, no data retained between executions

SCOPE

Synapse is not a general-purpose operating system. The production runtime runs on Linux via wasmtime. The UEFI bare-metal kernel in boot/ is a research component — it demonstrates the architecture can eliminate the OS entirely, but it is not the production path today.

The Python support is a constrained subset (arithmetic, conditionals, loops, functions, lists). It is not a general CPython replacement.

WHO NEEDS THIS

Verified execution maps cleanly to compliance and safety problems in regulated environments. The current product is a self-hosted runtime; broader domain-specific assurance remains part of the R&D stack.

Banking & Financial Services

OSFI B-13 · SR 11-7

AI loan approvals, risk calculations, and compliance reporting must be explainable, reproducible, and auditable. Regulators don’t accept “the model said so.”

Current proof path: deterministic execution, receipts, and a bounded verification boundary for generated logic. Broader privacy and compliance guarantees are active research, not the default product claim.

Healthcare & Medical Devices

Health Canada · FDA

Diagnostic AI must never return values outside calibrated ranges. Proving “never returns dangerous values” requires more than statistical testing.

Relevant R&D: output bounds, self-gating inference, and verified fine-tuning are demonstrated research threads. The current stable gateway boundary remains limited to @inv pure, @inv terminates, and @inv no_oob.

Government & Defense (Sovereign AI)

Data Sovereignty

AI infrastructure depends entirely on foreign vendors (NVIDIA, Microsoft, Google). Data sovereignty laws require domestic control of sensitive computation.

Current product:self-hosted execution on commodity hardware with no guest filesystem, no guest network, and deterministic receipts. Deeper sovereignty work such as bare-metal and networked compute remains R&D.