aiqbench analysis
This paper introduces "falsifiable release gates" for self-improving AI agents, replacing subjective safety claims with machine-checked invariants that persist as capabilities expand from 20% to 70% accuracy on real hardware. By enforcing a fixed set of standing rules across six releases and adding complex features like provable unlearning without breaking core guarantees, the authors demonstrate a scalable framework for verifying autonomous system behavior rather than relying solely on post-hoc guardrails.
Falsifiable Release Gates for Self-Improving Systems: Standing Invariants at Scale
arXiv:2607.13070v2 Announce Type: replace-cross Abstract: Safety claims for self-improving agent runtimes are almost always self-graded: a policy file, a guardrail, a promise in a README. We describe falsifiable release gates, a methodology in which every new capability must pass a pre-declared, machine-checkable acceptance suite before it ships, while a fixed set of standing invariants is preserved across every gate. We instantiate it in Antahkarana, an open runtime, then do what a method paper is only vindicated by: we follow the same runtime as it grows and ask whether the guarantees survive. The safety-critical property, that no action reaches an effector without a capability token minted by a control ring, is machine-checked exhaustively over the reachable states of a bounded model; a deliberately broken model yields the shortest counterexample, so the checker demonstrably has teeth. We then carry the runtime through six further releases. Across every one, the action-safety invariants INV-1 through INV-6 held without a single change, and one release added three capabilities while introducing no new invariant. Under the same teeth discipline, six more machine-checked families were added: memory with provable unlearning, a governed agent, calibrated abstention over a post-quantum record, a harness of many sub-agents, the self-improvement loop itself, and the residency of what it produces. The acceptance suite grew from 122 tests to 563. The load-bearing result sits in the negative space: across more than a doubling of capability, the safety core was neither weakened nor redesigned. The last families are the first on real hardware: gated self-improvement compounds a small model from 20% to 70% accuracy while auto-rejecting a candidate that only inflates confidence, and the whole governed path costs 0.021 ms per request, 0.008% of model inference. We release the runtime, tools, and gate suite; every number reproduces with a single command.
Abstract:Safety claims for self-improving agent runtimes are almost always self-graded: a policy file, a guardrail, a promise in a README. We describe falsifiable release gates, a methodology in which every new capability must pass a pre-declared, machine-checkable acceptance suite before it ships, while a fixed set of standing invariants is preserved across every gate. We instantiate it in Antahkarana, an open runtime, then do what a method paper is only vindicated by: we follow the same runtime as it grows and ask whether the guarantees survive. The safety-critical property, that no action reaches an effector without a capability token minted by a control ring, is machine-checked exhaustively over the reachable states of a bounded model; a deliberately broken model yields the shortest counterexample, so the checker demonstrably has teeth. We then carry the runtime through six further releases. Across every one, the action-safety invariants INV-1 through INV-6 held without a single change, and one release added three capabilities while introducing no new invariant. Under the same teeth discipline, six more machine-checked families were added: memory with provable unlearning, a governed agent, calibrated abstention over a post-quantum record, a harness of many sub-agents, the self-improvement loop itself, and the residency of what it produces. The acceptance suite grew from 122 tests to 563. The load-bearing result sits in the negative space: across more than a doubling of capability, the safety core was neither weakened nor redesigned. The last families are the first on real hardware: gated self-improvement compounds a small model from 20% to 70% accuracy while auto-rejecting a candidate that only inflates confidence, and the whole governed path costs 0.021 ms per request, 0.008% of model inference. We release the runtime, tools, and gate suite; every number reproduces with a single command.
23 pages, 14 figures. Major revision merging the follow-up "Standing Invariants at Scale" into this paper per arXiv moderation: the machine-checked action-safety core preserved across six further releases, six new invariant families with teeth, and real-hardware self-improvement; suite grown from 122 to 563 tests. Software, gate suite, run artifacts, and TLA+ spec are open source
Software Engineering (cs.SE); Artificial Intelligence (cs.AI)
D.2.4; D.2.5; I.2.0
arXiv:2607.13070 [cs.SE]
(or arXiv:2607.13070v2 [cs.SE] for this version)
Full text under the publisher's open licence · 364 words
Related coverage
AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connect
NVIDIA · Sep 29, 2026
Links #5: 2026/07
LessWrong · Jul 31, 2026
Why do models task game?
AI Alignment Forum · Aug 6, 2026
[AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M
Latent Space · Sep 22, 2026