Mo is a strict functional language whose OCaml compiler generates efficient imperative Zig — no ownership annotations, no garbage collector. Linearity inference, stateful pattern fusion, and opaque linear handles let pure functional code match or beat hand-written Zig on memory across 5 benchmarks.

Sparse attention trained end-to-end absorbs the routing signal into Q/K/V, so learned gates barely beat random ones ('routing absorption'). Modal experiment code and the paper showing post-hoc KL distillation sidesteps it — near-oracle quality at 87.5% sparsity on Qwen3-1.7B.

Minimum state products on Dijkstra's self-stabilizing token ring — exact values for n = 3 through 9 and a large-n conjecture supported through n = 10. Paper, sorry-free Lean 4 formalization, Python research probes, and an independent verifier.

Hamiltonian decomposition of Z_m^3 Cayley digraphs for all m > 2 — a complete solution to Knuth's "Claude's Cycles," found by LLM agents under a structured exploration prompt. Constructions, proofs, Lean formalization, and a verification suite.

Homebrew tap for the Cauldron suite of encrypted, peer-to-peer command-line tools: seance (ephemeral P2P chat), mitt (encrypted file transfer), omen (anonymous voting), covenant (membership signing), and familiar (chat-bot daemon).

Encrypted CLI tools written in Zig. No accounts, no cloud, no history.

Orthogonalizing an mLSTM's memory read is a removable training scaffold, not a memory gain — paper, code, and per-seed results.

Sorry-free Lean 4 proof that there are C(d·m, d−1)/d degree-d critical portraits, for all d.