ARKHE: A Verifiable Trust Architecture for Autonomous AI Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Autonomous AI agents are beginning to execute consequential actions — payments, deployments,signatures — on the basis of representations their principals cannot independently verify. Thismonograph develops and examines a design answer: Arkhe, a verifiable trust layer whose load-bearingelements are a fail-closed kernel of containment, cryptographic model attestation (canonicalized on theGGUF open-weight container), an append-only causal evidence graph (WormGraph) anchored toexternal transparency logs, and quorum-governed rule corpora (constitution-as-code) whose authorityis structurally separated from event witnesses. The study proceeds by design-science analysis withsystematic documentary verification. Incident reconstructions of the Bybit (2025, ~$1.5B), StablR(2026, $13.5M), and SUPERFORTUNE (2026, ~$15.18M) failures are mapped to architecturalcontrols; the architecture is aligned with the EU AI Act, GDPR, Brazil's LGPD, MiCA, and ISO/IEC42001; and a minimal viable product — GGUF model verification for machine-learning engineers — isspecified with falsifiable adoption metrics. Extending the architecture to its limiting case, the bookspecifies a five-front evidence program for the Hodge conjecture, a Clay Millennium Prize problem,integrating verified 2025-2026 results (including Markman's proof for abelian fourfolds) with Lean 4formalization and Rekor anchoring, and recording three displayed corrections to the program's internalclaims. Counterarguments, limitations, and open problems are treated as load-bearing sections. Theconclusion assesses the project as architecturally coherent but pre-product, and prescribes ametrics-gated path from verification artifact to market.