resolvent
A JEPA planner ranks candidate futures one at a time. When they share one error, the right pick depends on the whole candidate set.
- Python
- Lean 4
teerth.blog
Long-form essays on the projects, with the maths, the proofs and the numbers.
The family
Four lines of work as a family tree, one lane each: what each project was built on or came out of, and the upstream pull requests it led to. Every line is a link the author has confirmed.
Relations in the family: topological-ml-toolkit is built on Aether-Lang. monodromy came out of topological-ml-toolkit. nerve came out of topological-ml-toolkit. separatrix came out of Epsilon-Hollow. tangle came out of Epsilon-Hollow. planimeter came out of Epsilon-Hollow. faraday is in the same line of work as resolvent. Electromagnetic-Field-Data-Simulator came out of faraday. Asmodeus came out of resolvent. Asmodeus came out of caustic. sigmoid came out of caustic. sigmoid came out of Epsilon-Hollow. topological-ml-toolkit, monodromy, nerve form one group, topological tools, under Aether-Lang. topological tools led to triton-lang/kernels pull request 22: topology-derived sparse attention, 3.48× at 4K context. topological tools led to google/XNNPACK pull request 10801: memory planner reuses leading gaps, FP32 MobileNet V1 peak allocation 6.42% lower. topological tools led to google/highway pull request 3244: PHAST build 1.65× to 1.98× faster at 1,000,000 keys, output digests identical. topological-ml-toolkit led to google-deepmind/mujoco pull request 3396: island discovery on a disjoint-set forest, 1,281× less solver memory. topological-ml-toolkit led to NVIDIA/topograph pull request 432: chart RBAC rules gated by engine and provider. nerve led to open2c/polychrom pull request 79: getLinkingNumber sign fix. Epsilon-Hollow led to openxla/xla pull request 46539: deterministic GPU reduction grouping, landed as 3d5df1d. Epsilon-Hollow led to tensorflow/tensorflow pull request 124410: exact transitive closure of collective control edges. Epsilon-Hollow led to NVIDIA/NeMo-Relay pull request 481: cross-process agreement 0.48 → 1.00. Epsilon-Hollow led to google-deepmind/mujoco pull request 3450: convex hull graph without the quadratic scan, 4.21× faster cold load at 40,962 vertices. Epsilon-Hollow led to google-deepmind/mujoco_warp pull request 1541: island discovery on a GPU disjoint-set forest, scratch memory down 97.73%. caustic led to facebook/pyrefly pull request 4180: test capturing a capped-recheck propagation panic, landed as 3e90baa.
open2c/
google-deepmind/
NVIDIA/
triton-lang/
google/
google/
openxla/
tensorflow/
NVIDIA/
google-deepmind/
google-deepmind/
facebook/Each project links to its essay, each upstream node to its pull request. Only links the author has confirmed are drawn.
Essays
One per project: the problem, why it is hard, the maths, the proofs, what was measured and what failed.
A JEPA planner ranks candidate futures one at a time. When they share one error, the right pick depends on the whole candidate set.
Measuring the shape of a trained model's activations as a state, and rolling that state forward with one ridge solve instead of a second network.
A model can hold a fact and still fail to reach it. When it fails, distinct entities collapse onto one answer, and the collapse proves errors.
From a photo of two cables: an integer linking-number certificate, or a refusal that names the crossing to re-shoot.
Tells you whether a top-k list came from your data or from the computer's rounding, and refuses, naming the pair, when it cannot tell.
An agent edits a drawing and reports the room count. planimeter reads the file and returns the count, or the coordinate where the file cannot say.
A research kernel that keeps its state as points on a sphere and returns each discrete answer with a certificate that rounding could not change it, or refuses.
A research language where persistent homology is a primitive, so a loop can stop when the Betti numbers of its data stop changing.
Persistent homology as ordinary ML features, with a certificate that decides when one barcode can stand in for a whole moving point cloud.
Can a local atomic descriptor see how polymer chains are linked and knotted? A Rust harness that counts what the descriptor merges.
Testing whether a map can be run backwards by comparing pairs of points, instead of checking its Jacobian one point at a time.
Counts where a cavity mode's electric nodal lines meet the wall and where its magnetic field curls inside, and tests the exact law between the two.
A Rust, Python and TypeScript toolkit that builds the Faraday tensor from sampled E and B fields and summarises the energy map as a graph.
A stream whose law changes, then comes back. Asmodeus keeps one small table per law and a softmax over them, so a returning law is recalled, not re-learned.