Paper

ENF Whitepaper

Embedded Neural Firmware (ENF) reframes device software as a sealed Neural BIOS- offline-only, immutable, and bound to silicon-and as a framework that specifies, builds, and proves it. ENF compiles a task-specific, quantized model into ROM/Flash and executes bare-metal via a single FSM with static memory (no heap/GC or recursion) and per-state WCET, yielding bit-for-bit reproducibility and tractable verification. Identity and provenance derive from a Physically Unclonable Function (PUF), eliminating stored secrets and gating actuation through measured boot. Powered by harvested energy and supercapacitors under explicit thresholds (V_on/V_safe/V_cut), ENF targets years-long autonomy while collapsing the remote attack surface (no IP stack, OTA, or telemetry). We present ENF's invariants and the ENF Framework-a locked, reproducible toolchain and a Conformance Pack (signed Manifest, firmware/model hashes, PUF-bind record, SBOM, test vectors)-and ground the approach in three cloud-free deployments with a clear threat and limitations posture (recall/replace, dataset governance). Compared with TinyML and Cloud/Edge-AI, ENF offers privacy by architecture and assurance through reproducible builds-evidence you can verify today.