New AI Safety Tool Could Make Robots More Reliable

A new verification paper presents a substrate that exposes errors in its own claims, not just model outputs. GBI-DCSE v3 falsified an architectural claim about its reported Fisher value and budget, showing the claim held only on a slice of the full input box. BoundaryBench v0.1 established a baseline: Qwen3-4B-Instruct-2507 completed 768 frozen executions, but 0% cleared the contract, with 369 failing parsing and 399 failing validation. The companion study evaluates three improvements: BeTaL-GBI v0.2 applies benchmark tuning with an LLM-in-the-loop over 2,218,750,380 grid points, separating format admission from conditional performance; after schema repair, a model-free feedback search achieves a 2.87% mean held-out target gap, outperforming non-feedback baselines of 13.61% and 11.46%. GBI v2 swaps static keys for a reference-independent witness state and policy, detecting all 116 injected severe contradictions, accepting all 99 clean records, and blocking hallucinator and evidence-forger surrogates with zero silent promotions. GBI-DCSE v3 maps 99 claims to machine-readable evidence, with 95 of 96 testable claims passing and 148 standalone checks executed without failure across 62 configurations. Under synthetic conditions, the paper reports GBI-DCSE is a selective, policy-versioned, self-auditing test and routing substrate.