AdaCore Demonstrates Trustworthy AI Development with GNAT Foundry

Summary

AdaCore presents GNAT Foundry: Intersection, a GitHub demonstration of AI-assisted development for high-integrity software. The project uses a four-way traffic intersection with protected left turns and demand-led pedestrian crossings as an understandable safety-related example, rather than as a real traffic-control implementation. Its central proposal is to combine an AI agent with deterministic, non-AI tools so that people can focus on the artifacts where judgment is essential instead of reviewing every AI output. The demo begins with a CONOPS, high- and low-level requirements, an architecture description, Ada implementation, simulation harness, and traceability matrix. AdaCore's toolchain supplies compile-time checks, SPARK Silver proofs of absence of runtime errors, SPARK Gold correctness proofs, requirements-based tests, and reported 100% MC/DC and statement coverage. The resulting HTML report identifies the CONOPS and selected derived requirements as the main human-review points, while presenting deterministic evidence for the rest of the chain. Users can run the baseline with the open-source Ada SPARK toolchain or GNAT Pro, inspect the graphical simulation, and generate the report. A scripted change request asks an AI agent to add pedestrian countdown timers, update the requirements, architecture, implementation, tests, proofs, and report. The workflow is designed to expose any additional items requiring human review. AdaCore says the approach supports both brownfield systems with an existing evidence base and future greenfield projects. The company clarifies that the demo does not demonstrate a particular certification process, and that certified projects would still require human review of changed requirements and test adequacy.