VAREK Information

VAREK — deterministic, pre-execution verification of AI agent actions. varek-lang.org · github.com/kwdoug63/varek · MIT VAREK verifies what an autonomous AI agent is about to do, and proves it complies with an explicit policy, before the action takes effect. Medicine does not deploy systems that are "usually right"; VAREK applies that standard to agent behavior. Every action resolves to one of three verdicts — SATISFIED, UNSATISFIED, or UNKNOWN — discharged by a satisfiability-modulo-theories decision procedure. The runtime is fail-closed: an action proceeds only on a provable SATISFIED, and UNKNOWN is never coerced into a pass. That is the difference between a verifier and a heuristic. The release line, v1.0 to v1.9.2: v1.0 — public launch of the formal-verification layer (MIT) v1.5 — the Warden: kernel-level enforcement via seccomp-unotify; sub-microsecond decisions v1.6 — pre-execution verification of the whole action-graph; an unsafe plan is refused before its first action runs v1.7 — verification bound to enforcement at the system boundary, one vertical stack v1.8 — cross-action data-flow verification; a secret source cannot reach a forbidden sink v1.9 — progress-safety: a load-time liveness proof certifying human-out-of-the-loop operation v1.9.1 — enforcement hardening: io_uring bypass closed; TOCTOU-safe file mediation (measured 510→0) v1.9.2 (current) — mediation completeness: the Warden inverts to a default-deny allowlist, closing the x32 ABI bypass and denying never-legitimate syscalls and unprivileged user namespaces by construction; validated under live enforcement by a conformance target Open source, MIT. Three provisional patent applications (patent-pending) cover the SMT decision-procedure layer, the Warden kernel architecture, and action-graph compositional policy decision. Developed by Sober Agentic Infrastructure, Inc.