What is “Controlled” AI and how it supports Software Verification and Validation

The economics of software verification and validation are shifting. As generative AI lowers the marginal cost of producing requirements, tests, and code, the dominant cost in safety-critical projects is migrating further toward review, traceability, coverage demonstration, and evidence preparation. For V&V professionals, the operative question is: under what conditions AI generated outputs can be accepted into a verification workflow that remains compliant with EN 50128, ISO 26262, DO-178C, IEC 61508, ect. as well as with the emerging guidance such as ISO/IEC TR 5469, ISO/PAS 8800, and the EASA AI roadmap.

This presentation reports how Developair Technologies delivered two industrial pilots representing this challenge in concrete settings, with the explicit goal of validating the new TRL7 prototype, NOVA, and demonstrate which engineering practices made AI-assisted V&V usable, auditable, and defendable in front of safety and quality stakeholders.
The first pilot addressed the validation of SIL2 software in a Hardware-in-the-Loop environment, in a context where the customer was scaling its V&V process from basic-integrity software to a higher criticality level. Manual generation of test cases and execution scripts had become a bottleneck. The pilot combined an advanced analysis of the requirements as a set, AI-assisted test-case generation, expert-driven selection of branches and coverage strategies, and automatic translation of test cases into a proprietary in-house scripting language for execution on the real HIL bench.

Key lessons concerned:

1. The steps of the human-in-the-loop are controlled through a multi-tool approach including symbolic AI and deterministic techniques ensuring discrete and auditable decisions

2. Generative AI’s scalability challenges and elevated costs can be diluted using complementary deterministic algorithms

3. There is a practical effort required to maintain traceability in a set of highly connected artifacts while demonstrating safety standards’ compliance.

The second pilot focused on legacy SIL2/SIL4 software previously certified under EN 50128, where the cost of future recertification was driven by outdated, fragmented, or missing artefacts. The work explored AI-assisted recovery and refinement of requirements from existing code and documentation, regeneration of unit tests, automatic generation of executable scripts in a CXX test framework, and execution against the legacy code to obtain coverage figures, including MC/DC.

Lessons here centred on:

1. Value added to the achievement of code coverage from boosted AI black box test generation

2. The generation of requirements and tracing to the software architecture creates opportunities and benefits in simplifying other test suites and scenarios (crediting)

Across both pilots, common findings emerge: the integration of controlled AI techniques, through multi-tool and human-in-the-loop approaches, into existing processes can reduce the engineering effort up to 50% while ensuring safety standards compliance and minimal disruption to the existing workflows; the artifacts generated by this controlled AI are relevant and auditable independently from the original input.

Speaker

Edgar Hernandez

Senior Embedded Software Engineer, Developair Technologies

Engineer with more than 5 years of experience developing embedded software, passionate about automation, clean coding and AI integration for embedded systems and software development.