본문 바로가기
  • Home

Numerical Verification of LLM-Generated Illuminance Calculation Code: A Case Study Using Synthetic IES Photometric Data

  • Industry Promotion Research
  • Abbr : IPR
  • 2026, 11(4), pp.155~163
  • DOI : 10.21186/IPR.2026.11.4.155
  • Publisher : Industrial Promotion Institute
  • Research Area : Interdisciplinary Studies > Interdisciplinary Research
  • Received : September 21, 2026
  • Accepted : October 7, 2026
  • Published : October 31, 2026

Hyun Mok Park 1,  Hyun-Seok Choi 1,  Woojin Cho 1,  Seok-Oh Bang 1,  Ki-Taek Kwon 1,  Sang-Wuk Shin 1,  Jin-Sung Rho 1,  Jong-Min Lim 1

1키엘연구원

Accredited

ABSTRACT

This study verifies the outputs of illuminance calculation code generated through LLM-based vibe coding using synthetic IES photometric data. GPT-5.6 Sol was used at its lowest reasoning setting under three request configurations: a single request for the full calculation, six sequential requests, and the same sequence with explicit calculation requirements. Each configuration was executed three times. Nine outputs from six synthetic distributions and 144 queries were evaluated against analytical references and values calculated for five assumed error types. A design error in IES-E assigned different intensities to the same nadir direction for different azimuths, and the seven affected rows were excluded from accuracy assessment and analyzed separately. Among the remaining 137 rows, the maximum relative error for queries targeting tabulated angles was below 1.0×10⁻⁵ %, while the mean and maximum errors for intermediate-angle queries were 0.1013 % and 0.3780 %, consistent with the error expected from linear interpolation of the 5° tabulation. No outputs matched the assumed-error values at points where those values could be distinguished from the analytical references. Differences between request configurations were limited to output recording and floating-point precision, while two of the nine runs differed only at five of the seven excluded rows. These results highlight the need to examine both numerical errors across photometric distributions and calculation angles and the physical consistency of test data when verifying LLM-generated illuminance calculation code.

Citation status

* References for papers published after 2025 are currently being built.