Every project below follows the same discipline: an idea, stated so it can fail; a measurement that could kill it; a manuscript written from programmatic macros so no number is retyped; and a public artifact trail — committed per-seed result JSONs, one-click reruns, trained checkpoints on Hugging Face, result mirrors on Kaggle — so every claim can be traced and re-executed. The common thread is structure: conservation laws held exactly (Metriplectic), geometry carried through depth (AGF-NO), governing equations in attention (PhysFormer), stability guaranteed by proof (certified Lyapunov control), and benchmarks with independent verifiers (PhysBench). Where an idea failed, the failure is reported in the same typeface as the wins.
• Adaptive Contact Dynamics • Age Artificial General Engineer • The Cost of Automatic Action Labelling • Verifying Neural Lyapunov Certificates • Certified Training of Lyapunov Functions • Compositional Grid Certificates • The Dirty Man: Switch Operators • Dreamerv3 Latent Stability • Electric_Adaption • Gnome • Gnome Manufacturing • Gnome Psn1 Nmi • Morphogenetic Bimodal Networks • Multi-Agent ISS Certificates • Operating Atlas • Operator • Perch • Honest Physics-ML Components • PhysRNet • The Rank-Fidelity Paradox • DataFly: Robot Data Curation • Self Modeling Stable Control • Spectral–Topological Decoupling • Steerable VLA • Stochastic Latent Bounds • Tactile Process Control • The Tolerance Law • Certifying the Residual Stream • Ugct • Verified Voltage Control • Website