AI governance infrastructure · formal methods · code-backed research.
Independent researcher and infrastructure builder. M.S. Computer Engineering, New York University. Based in Yokohama.
Which Structural Constraints Are Learnable? A Regime Map for a Minecraft Voxel Generator — Accepted at the FDG 2026 17th PCG Workshop
Code-backed study of structural learnability in 32³ Minecraft voxel generation. VQ-VAE + AR transformer over ~260 block types; 14 structural properties → 9 controllable, 4 approachable, 1 unresponsive. Composite predictor signal × min(CV, 1) reaches Spearman ρ = 0.879 (p = 0.002, n = 10). Public DOI and reproducibility package are available below.
Blackwell Dilemma: Information Value Under Endogenous Feasibility
Information value can be non-monotone in signal informativeness when feasibility is endogenous. A five-state model exhibits an explicit phase transition; classical Blackwell dominance reverses under generic feasibility curvature.
Accountability Arbitrage: Ethical Tensions in AI Agent Accountability Infrastructure — Under review at AI and Ethics
When firms can choose where AI agent actions are reviewed, accountability becomes arbitrageable. Maps power asymmetries in the AEGIS protocol's three-stage review pipeline; identifies labor-market spillovers when downstream verification migrates across jurisdictions.
Full research list, including SSRN-visible working papers in AI governance, decision theory/economics, algorithmic foundations, and philosophy of cognition, at crabsatellite.com.
AEGIS Protocol — Open protocol for AI agent action accountability. Apache 2.0, 202 tests, zero dependencies. Reference implementation of the protocol behind the Accountability Arbitrage paper.
Dockerfile Doctor — The only Dockerfile linter that fixes what it finds. 80 rules, 51 auto-fixers, pure Python, zero dependencies.
Minecraft Modding — Independent mod author on CurseForge. Total downloads across maintained projects: 1.49M+. Ported 9 mods to NeoForge/Forge 1.21.
