Child1

child-1 makes LLM decode faster on the Qualcomm Hexagon NPU and proves every speedup on a real phone. It optimizes llama.cpp's Hexagon NPU backend (ggml-hexagon: HVX and HMX kernels, DMA, VTCM), where phone NPU inference speed is won or lost. child-1 is three Fetch.ai agents on the Agent Chat Protocol, used entirely from ASI:One: child-1 (orchestrator), NPU Verifier (judge) and NPU Benchmarks (measured NPU speeds). Behind them, a self-improving loop runs sandboxed LLM coding workers (Muse Wonder, Gemini 4 Argon, Claude) that write NPU kernel optimizations and may only edit ggml/src/ggml-hexagon/. A locked verifier with no LLM in it builds each patch, then races baseline and patch in alternating rounds on a real Galaxy S26 NPU (Snapdragon, Hexagon V81). A patch passes only if every layer runs on the NPU with no CPU fallback, the output hash is identical, decode tok/s beats measured noise, and prefill, time to first token and memory do not regress. Failures become versioned changes to the system's own NPU instructions, and a new version goes live only if it beats the old one on a model the agents never saw. SubQ Snap groups NPU failures for the supervisor, and Muse Spark summarizes lessons. LLMs write NPU kernels, the phone's NPU decides, and the agents can't edit the judge.