Projects

Pick one and the seal slides there.

What I fixed, and where. 12

Eleven landed contributions. Every number below was measured on a named machine against a control, and every one of them sits in a merged diff or on the default branch it targets.

• Quadratic scratch in island discoverygoogle-deepmind/mujoco #33961,281.6x less scratch
• A linear-memory GPU disjoint-set uniongoogle-deepmind/mujoco_warp #15411.513x faster
• An inverted point-id table for convex hullsgoogle-deepmind/mujoco #345015,361x fewer probes
• Pruned collisions by slice structuregoogle/highway #324465.5x fewer comparisons
• Reusing the leading gap in the memory plannergoogle/XNNPACK #108016.42% lower peak
• A missing back-propagation in transitive reductiontensorflow/tensorflow #1244104 edges to 3
• One scaffold, not one profile per taskNVIDIA/NeMo-Relay #48123 files
• A topology-derived sparse attention kerneltriton-lang/kernels #22804 lines added
• Deterministic GPU codegen in reduction groupingopenxla/xla #465395 lines, deterministic
• Gating cluster wide RBAC to the rules that applydsx-ai-factory/topograph #432145 lines gated
• Pinning a 208-module recheck propagation panicfacebook/pyrefly #4180208 SCCs pinned
• Fix the sign of getLinkingNumberopen2c/polychrom #7911/11 pairs agree

In the lab 11

• resolventAttention and Markov paths share one operator.
• Epsilon-HollowMemory, files and scheduler, on one sphere.
• Aether-LangLoops stop when their shape stops changing.
• causticHallucination, measurable with no ground truth.
• monodromyCan it be undone? Topology answers.
• topological-ml-toolkitThe shape of data, as an ordinary feature.
• faradayThe field coupling, found rather than assumed.
• nerveThe control that could kill my result.
• separatrixDecided by the data, not by rounding.
• planimeterExact, or refused.
• tangleIt refuses rather than guesses.