sheaf-core

Core algorithm for Sheaf-ADMM multi-agent coordination, in pure Rust (ndarray only): agent graph with directional slot tables, cellular-sheaf geometry (fixed and LoRA-factored restriction maps, matrix-free sheaf Laplacian), closed-form diagonal-prox x-solvers, a batched unrolled conjugate-gradient z-solver (project and prox modes), and the unrolled ADMM loop with per-iteration history (primal/dual residuals, consistency RMS).

This is a parity-faithful port of the forward path of SakanaAI/sheaf-admm (JAX/Flax, Learning Multi-Agent Coordination via Sheaf-ADMM, ICML 2026, arXiv:2605.31005) — verified layer-by-layer and per-iteration against golden traces from the reference implementation.

Part of the sheaf-rs workspace: sheaf-nn provides the neural encoders/decoders and full model, sheaf-io provides weight/fixture loading and data views.

Licensed Apache-2.0. Derived from sheaf-admm (Apache-2.0); see the repository NOTICE for attribution.