Sovereignty infrastructure. BAION Systems builds instruments that give people evidence of what is happening to them — receipt-bearing, reproducible, and owned by the person who generated it, not the service they happened to use.

Evidence was the first half. The same instruments now run forward: coordination that keeps a schedule — or a team of local AI agents — moving with no central dispatcher to lose, and a receipt for every decision it made. The coordination substrate is an active research program: agents couple through a shared dynamical medium instead of waiting on a central orchestrator, and we are testing whether that coupling can support consensus, task allocation, and recovery from disruption — claims bounded by what runs prove.

A service-disabled-veteran-owned small business in Raleigh, North Carolina. Founder-led, local-first, receipts-driven: every experimental claim in our work is backed by hash-chained run records on hardware we own.

• baion-std — seven languages (C, C++, Rust, Go, D, Haskell, OCaml) producing byte-identical canonical JSON + SHA-256, with a one-command cross-lineage verifier, re-run in CI on every push to main and every pull request — checking byte-identical agreement across all seven lineages over the tested corpus and differential probes (tested agreement on the covered inputs, not a proof of canonicalization correctness against an independent oracle). The trust discipline underneath our multi-component systems.
• baion-vault-method — the governance method behind the work: receipts establish state, exact-byte review and promotion, positive controls before negative results, preserved retractions — as a reference methodology with executable enforcement, not just documentation.
• synchronization-resistance — a pre-registered study measuring how hard it is to get independent LLMs to agree. Three hypotheses passed, one was inconclusive, one failed — all reported. Pre-registration and archival DOI linked from the repo; not peer reviewed.
• framework — essays on multi-model consensus, accountability, and why convergence is the product. Historical material from early 2026; later work bounds some of its claims, and the repo carries editor's notes saying where.
• More at baion.dev — instruments, evidence (DOI, pre-registration, federal filings), and method.
• Live demo: demo.baion.dev — the Shadow Comparator on the real engine. Practice data, clearly badged; every run carries a replayable receipt.

The coordination substrate now has two private, invitation-only evaluation lineages: a benchmark-neutral package and a separate MAFBench adapter package. Both carry runnable examples, contract tests, architecture notes, and reproducibility receipts. They are evaluation surfaces, not published benchmark results, and no Dairy Bench integration is claimed.

The underlying mechanisms are described across U.S. Provisional Patent Applications #64/042,046 (filed April 17, 2026), #64/143,042 (filed August 28, 2026), and #64/145,679 (filed September 1, 2026). These are provisional applications; no patent grant is asserted. The engine source and full internal harness remain unreleased pending counsel review. BAION's long-run model is open: the business charges for responsibility, not control.

Steven A. Mullins Jr., Founder / Managing Member — email preferred.