Architecture research exploring capability-tiered AI governance and compute-linked enforcement protocols.
Interactive signaling game on compute governance, U.S.-China AI competition, and Vietnam as a downstream observer.
Modeling when labs comply, evade, or split compute across jurisdictions and what drives each outcome
Model, data and verification suite for "The Half-Life of Compute: Modelling National AI Capacity Decay Under Supply Severance" (Escoda, 2026). Reproducible: fixed seed, independent re-derivation of every headline number. EN/IT/FR.
Detection-probability simulator for recomputation-based covert-training detection -- the verification mechanism in AI 2040: Plan A.
Game theory of the AI race: why the OpenAI–Anthropic–China competition is a prisoner's dilemma, chicken or stag hunt depends on how one lab's carelessness adds to everyone's risk. Working paper, replication code, tests Amodei's pacing proposal.
An open protocol for recording why AI workloads are placed, routed, migrated, and executed across distributed compute resources.
Open, versioned, publicly-contestable risk scoring format for AI verification coverage and frontier-system risk -- CVSS for AI safety.
Reproducible tools for recipe discovery and compute-governed experiment promotion.
Working implementation of pieces of the AI Futures Project's AI 2040: Plan A verification architecture — real trained model, real DiFR server, real red-team attacks. 30/30 sections drafted.
Verifies Ed25519-signed compute-attestation records and separates verified from unverified totals. CLI + MCP server.
米中AI競争の多層構造 ── 決めているのは、強さではなく条件である / The Layers of US-China AI Competition — What is being decided is not who is stronger, but who writes the terms.
To associate your repository with the compute-governance topic, visit your repo's landing page and select "manage topics."