Deepak Soni's falsifiable gates ensure self-improving AI safety with formal verification

Deepak Soni's paper, "Falsifiable Release Gates for Self-Improving Systems," tackles a core AI safety challenge: how to trust agents that rewrite their own code. Instead of relying on self-graded policies or guardrails, Soni proposes a formal methodology where every new capability must pass a pre-specified, machine-verifiable acceptance suite before shipping, with fixed invariants preserved across gates. The method is implemented in the Antahkarana open runtime via seven gates, from basic observability to a self-governing loop that suggests policy changes. No action reaches an effector without a safety-critical capability token minted by a control ring, which is exhaustively machine-checked over the one million recorded reachable state space of a bounded model and re-checked against execution traces.

The system ensures positive constraint: the entire write surface consists of policy rules, with tightening changes auto-applied and loosening changes always requiring human merge. A proposal that mispredicts its own effect is auto-closed. Soni publishes measured acceptance results for all seven gates, precisely scoping claims to the coordination skeleton (not learned components). The runtime, command-line tools, gate suite, TLA+ specification, and run artifacts are all open source, enabling reproducibility across other agent frameworks. Reviewers can run the central non-bypass with a single command in seconds. This represents a practical step toward verifiable self-improving AI, blending formal methods with real-world deployment.