3 minute read

On July 11th, 2026, the Natural Sciences and Engineering Research Council of Canada (NSERC) announced that the research program titled Safety-inspired language for non-safety-critical systems is awarded $210,000 as part of the Discovery Grant program. The program started on April 1st, 2026 and runs through 2031.

Safety engineering is a mature discipline, but its methods have always been reserved for systems whose failures are obviously catastrophic: aircraft, medical devices, nuclear control. Everyday software was never held to that bar, and never needed to be. That assumption is now breaking. A wrong recommendation in a therapy app is not a rounding error, and the developers writing it have no structured way to reason about safety at all, let alone demonstrate it.

The first of the program’s three objectives is jPipe’s research agenda stated in full: a safety-inspired language in which developers write down why a system is safe to deploy, a compiler that checks those arguments stay sound, and a composition algebra for building large arguments out of small ones. The “safety cards” mentioned under that objective are a thread already being pulled: at ICSE-NIER’26 model cards were reframed as safety artifacts, and Safety First!, to be presented next month at MoDRE, runs the idea at full size on the GPT-5 system card, tracing 74 published claims back to 22 requirements nobody had written down. The two remaining objectives push outward: organizing reusable safety knowledge so developers can find it, then wiring the whole thing into ordinary development so that a change preserving behaviour also preserves safety.

This grant follows the Large Scale Composition Discovery Grant (2020 - 2026), whose work on composition operators became one of the keystones of jPipe.

Abstract

This research program pioneers the study of safety for non-safety-critical (non-SC) systems, focusing on everyday software rather than high-risk domains like aerospace or nuclear control. While classical Safety-Critical (SC) systems follow strict regulatory frameworks, non-SC software increasingly carries safety implications, especially as AI-powered applications expand into domains such as mental health. A wrong recommendation or privacy breach in a therapy app can cause genuine harm. Yet, current development practices provide no structured way for regular developers to reason about or demonstrate safety.

This program aims to close that gap by transferring principles from system safety engineering into practical, developer-friendly methods. Over the next five years, the research will deliver three short-term objectives (STOs): (i) to design a safety-inspired programming language enabling developers to express why their systems are safe to deploy (STO1); (ii) to categorize and organize reusable safety assets that guide decision-making (STO2); and (iii) to integrate these mechanisms into standard software lifecycles to support qualification and potential certification (STO3).

STO1 will create a lightweight domain-specific language, supported by a compiler that performs syntactic and semantic checks to verify safety-related claims. Developers will be able to express arguments such as “the feedback system is rule-based and psychologist-approved,” and the compiler will ensure that such claims remain logically and semantically sound in the developed software. A composition algebra will let developers build complex safety arguments from smaller sub-claims, while “safety cards” (adapted from AI model cards) will enable empirical validation against real-world systems and repositories such as GitHub and Hugging Face.

STO2 will structure the diverse landscape of safety-related decisions by linking them to their underlying goals. Using goal modelling and product family engineering, it will identify universal versus domain-specific practices, such as post-processing safeguards (generic) versus text sensitivity in mental health apps (domain-dependent). The resulting framework will help developers detect risks, propose mitigations, and avoid unsafe design choices, such as misusing generative AI for therapeutic guidance.

STO3 will embed these ideas into the developer workflow, connecting safety claims directly to code and ensuring that any modification that preserves functionality also preserves safety. It will culminate in a qualification process that adapts principles from standards like ISO 26262 to non-SC software, paving the way for lightweight certification of non-SC systems.

The program’s impact extends well beyond its initial focus on AI and mental health. By providing open-source tools, open-access publications, and direct engagement with practitioner communities, it aims to reshape how developers reason about safety.

Consortium

  • McMaster Centre for Software Certification (McSCert)
    • Sébastien Mosser

Updated: