Skip to Main Content

Logan W.

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Logan W. — Demo video

muse-softfloat is a clean-room IEEE 754-2019 software floating-point library for the RP2040, a microcontroller with no FPU: binary16, binary32 and binary64, 849 operation instantiations covering the full required operation set plus 66 recommended ones, MIT licensed, built from the standard’s text alone with no third-party softfloat source. The proof that it works: REBOUND’s IAS15 N-body integrator runs unmodified on a Raspberry Pi Pico through this library and produces bit-identical state to the desktop build, 2,001 of 2,001 step records, on each of six boards. KissFFT and an interval-geometry catalogue match the same way. Every operation is checked against 411,213 committed test vectors (exact arithmetic, cross-checked against MPFR and Berkeley TestFloat at generation), CI runs the gates on Linux and Windows, and the library runs live in the browser as WebAssembly at loganw234.github.io/muse-softfloat. The second deliverable is the build itself as an experiment: a single Muse Wonder session led the whole project under two contract-driven working methods (HonestFramework and ParcelRound), dispatching 128 recorded Muse agent sessions as authors, verifiers and auditors, with an independent verifier checking every parcel against a copy carrying planted defects before merge (187 of 195 caught). I typed an estimated 2,000 tokens against 1.42 billion billed. The dated record, metrics ledgers and a frozen, hash-manifested audit bundle ship in the repo.

Team