Reliability Assurance for LLM-Assisted Energy-Control Software: Deterministic Verification and Reproducible Fault Injection for Battery Peak-Load Compensationtionsation

Reliable energy-control software must convert telemetry and battery constraints into reproducible commands. Large language models can assist development, but their outputs require verification before deployment. Reliability assurance combines lifecycle measures and supporting evidence; this experiment evaluates its automated verification layer, not operational reliability. A deterministic battery peak-shaving controller enforces power, energy, efficiency, state-of-charge, state-of-health and ramp constraints, including an immediate zero-excess reset. The reference implementation passed thirteen controller tests and six output-contract tests. Three stipulated low-, medium-, and high-variability fault profiles generated 9 000 candidate responses for three scenarios, with 1 000 runs per profile–scenario combination. Each candidate was assessed for branch selection, aggregate threshold, discharge setpoint, mean absolute percentage error, output-schema compliance, derived-field consistency and physical safety. Primary error proportions were 6.0–6.4%, 15.9–16.7% and 29.7–30.2%; overall pass rates were 72.2–73.4%, 38.7–43.5%, and 15.9–16.4%, respectively. Schema-only validation accepted control-invalid responses in 13.9%, 32.1% and 45.8% of all runs and safety-violating responses in 3.9%, 9.9%, and 17.0%. Full semantic verification admitted neither class in the tested corpus. These results characterize verifications under the declared synthetic mutation model; they neither estimate commercial language-model failure rates nor establish universal detection completeness. The proposed architecture retains real-time authority in deterministic code, with generated development artefacts subject to automated verification, independent review, and signed release.