Safety claims on self-improving agent runtimes are almost always self-graded, such as through policy files, guardrails, or README commitments. We describe falsifiable release gates and a methodology to build and validate such systems, ensuring that every new capability must pass a pre-specified, machine-verifiable acceptance suite before it is released, while preserving a fixed set of standing invariants at each gate. We applied this method in Antahkarana, an open runtime, via seven gates, ranging from basic observability to a self-governing loop that suggests changes to its own policy. No action can proceed to an effector without a safety-critical property capability token minted by a control ring, which is exhaustively machine-checked over the one million recorded reachable state space and re-checked against execution traces. A purposely broken model provides the shortest counterexample, ensuring the checker is effective. The self-enhancement loop is positively constrained: the entire write surface consists of policy rules, tightening changes may auto-apply, while loosening changes always require human merge. A proposal that mispredicts its own effect is automatically closed. We publish the acceptance measured results for all seven gates, precisely defining the scope of each claim (bounded to the coordination skeleton, not the learned components) and freeing the runtime, tools, and gate suite for reproducibility, allowing gates to run against other agent frameworks. Reviewers can repeat the single command central non-bypass in seconds.
Blogger's Review: This paper introduces a novel falsifiable release gate mechanism that enhances the safety of self-improving systems, particularly in scenarios requiring strict change control. The use of machine-verified acceptance tests and invariance checks significantly boosts the reliability and controllability of the system, especially in complex self-improvement environments, warranting further exploration and application.