AI agents are lowering the cost of formal methods across the board. They can help reason about existing code, propose invariants, write annotations, generate tests, and find counterexamples. That is real progress, especially for legacy systems where verifying code in place is often the only realistic path. But cheaper retrofit verification is still retrofit verification: the proof must contend with host-language semantics, existing APIs, mutation patterns, and performance tricks that were not designed around formal reasoning. For new high-assurance components with clear boundaries, I argue that certified programming remains the cleaner style: write the specification, implementation, and proof together in a proof assistant, then extract the certified program into the production language. This talk presents Crane, Bloomberg’s extraction system from Rocq to modern C++, as a case study in making that route practical.
Crane focuses on the deployment side of certified programming. Its goal is not merely to emit C++ that compiles, but to generate code that production engineers can read, review, test, debug, and maintain. Crane maps Rocq programs onto modern C++ abstractions: inductive types become variant-based data structures, polymorphism becomes templates, type-class-like interfaces become concepts, and recursive functional programs can be transformed into explicit loops to avoid relying on the C++ call stack. Crane also supports custom mappings from Rocq definitions to native C++ libraries, so certified components can use the abstractions already accepted in the host codebase.
I will discuss the engineering trade-offs behind this design: readability versus performance, functional source programs versus idiomatic C++, proof-oriented abstractions versus existing library interfaces, and assurance versus deployability. I will also give an honest account of the trust boundary: Crane itself is a lightweight, unverified extractor. Even though the Rocq program is checked against the Rocq specification, confidence in the generated C++ comes from readable output, conventional testing, static analysis, AI-assisted bug finding, and differential testing against other Rocq compilation paths.
The broader claim is that industrial formal methods need a portfolio, but the choice of programming style still matters. AI makes both retrofit verification and certified programming more affordable. Crane is built on the bet that, when we can choose the component boundary, correct-by-construction programming gives us a better engineering foundation. The remaining challenge is the last mile: making certified components look and behave like ordinary modern C++ libraries.
I am a researcher at Bloomberg’s CTO Infrastructure and Security Research team in New York City, applying formal methods in industry.
Wed 7 OctDisplayed time zone: Pacific Time (US & Canada) change
15:30 - 17:00 | REBASE 9SPLASH/ISSTA REBASE industry track at Grand Ballroom Salons A+B Chair(s): Daniel Lehmann Google, Germany | ||
15:30 45mIndustry talk | What Agentic Development Means for Rust (and Vice Versa)invited-talk SPLASH/ISSTA REBASE industry track | ||
16:15 30mIndustry talk | The Last Mile of Certified Programming SPLASH/ISSTA REBASE industry track Joomy Korkut Bloomberg Media Attached | ||