The work, ranked roughly by how much it matters and how far along it is. Each one links to the best version available today — a project site or a preprint — and will move to arXiv as papers are accepted.
A geometrically compiled language where logical operations over vector spaces are resolved at compile time into matrix multiplications. Branches become continuous weighted blends, loops become geometric rotations, and a whole program collapses to straight-line tensor work — GPU-native and differentiable by construction. The language paper and its NeurIPS 2026 supplementary are published alongside the project.
Relational displacement analysis on frozen text-embedding models, using Wikidata triples as probes. Two results: 30 model-agnostic relational operations that encode as consistent vector displacements, and a silent [UNK] tokenizer defect in mxbai-embed-large where unrelated diacritical strings collapse to cosine 1.0. Paper ID 2604.00648.
A lean, high-performance RDF-star triplestore in Rust with native HNSW vector indexing, ontochronological temporal queries, and a SPARQL+ query language — one engine where vectors are just triples, so a single query traverses relationships and similarity together instead of bolting a vector database onto a graph store. Systems writeup and benchmarks live in the repo.
Does redemption-narrative steering measurably move emergently misaligned LLMs back toward alignment — behaviorally, on self-rated harmfulness, and geometrically against a derived misalignment direction? Theoretical framing: emergent misalignment as moral injury. Cross-scale results across Llama and Qwen (0.5B–8B); paper and synthesis live in-repo.
A neuro-symbolic, GPU-native operating system written in Sutra: kernel, processes, IPC, and GUI as one differentiable tensor-op graph, with processes exchanging structured embeddings instead of bytes. The architecture and the argument for it are written up in the repo.