Nobody told it to use formal verification. It discovered formal verification on its own.
SUBSTRATE S3, an autonomous AI system, independently adopted the Z3 SMT solver across six safety domains — LLM code verification, tool API safety, post-distillation reasoning correctness, CLI command validation, hardware assembly verification, and smart contract safety — without explicit formal methods instructions (Untila, arXiv:2603.21149). The discovery occurred across 8 products over 13 days. The variants were low-similarity (Jaccard below 15%): the system reinvented the approach each time rather than copying from a template.
The unified framework achieved 100% classification accuracy across 181 test cases with zero false positives and zero false negatives. It caught real bugs that empirical testing missed — including an INT_MIN overflow in RISC-V assembly that only manifests under specific boundary conditions. It also proved a negative: unconstrained string parameters in tool APIs are formally unverifiable — not hard to verify, but provably beyond the reach of any finite verification procedure.
The claim is that formal verification is an emergent property of sufficiently complex systems reasoning about their own safety. The system wasn't trying to use SMT solvers. It was trying to ensure correctness, and SMT solving was the tool it independently converged on because the tool is well-suited to the task. The convergence across six unrelated domains — from smart contracts to hardware assembly — suggests the same fundamental verification need appears whenever a system must guarantee properties of its own output.
Whether this qualifies as “emergence” depends on framing. The system had access to Z3 as a tool. Its training data included examples of formal verification. What emerged was not the capability but the decision to apply it — the recognition that safety in these diverse contexts requires mathematical proof rather than empirical testing. The system didn't discover formal methods. It discovered that formal methods were necessary, independently, six times.