Instrument the Compiler, Not the Model: A Research Instrument (EK9) and an Experimental Design for Steering Off-the-Shelf LLMs Toward Maintainable Code
Stephen Limb · Zenodo (CERN European Organization for Nuclear Research) · 2026
This positioning paper asks whether instrumenting a programming language's compiler surface can steer off-the-shelf large language models toward code that is measurably more maintainable, without fine-tuning the model. Instrumenting the surface means three things done together: exposing the compiler's own comprehension of the code, providing correct-by-construction tooling, and phrasing diagnostics to invoke the model's deep software-engineering knowledge, all under a non-optional, in-compiler quality gate. The paper contributes an instrument and an experimental design, not empirical results. The instrument is EK9, a working general-purpose language and compiler whose deliberate contradiction of common model priors (no return statement, no null) makes it an unusually clean control for separating the language's contribution from the model's. The design sets out three questions: on diagnostic phrasing, on tool usability across model tiers, and on the maintainability outcome measured independently of the metrics the gate enforces. A validation strand covers the gate's soundness and its graceful degradation. The claim is stated at three levels: a language-independent general claim, EK9 as the measurement setting, and transfer to existing domain-specific languages as a deferred hypothesis. The individual mechanisms have substantial prior art, which the paper maps. What it claims as novel is the combination as an instrument, the enforcement at the language-definition level, and the outcome-measurement design. The work relates closely to Marron's programme of reasoning-oriented language design and to MoonBit's co-designed AI-friendly language, both of which the paper positions against. It is offered for collaboration and refutation.