Verifier-guided knowledge distillation of large language models for auditable thermal comfort reasoning

Occupant-centric building operation requires models that can not only predict thermal sensation, but also justify predictions through auditable reasoning. Traditional mechanistic models, such as Fanger’s predicted mean vote (PMV), often show limited field accuracy, whereas advanced machine learning models can improve prediction but typically remain opaque to facility managers and building operators. This study proposes a verifier-guided knowledge distillation framework for auditable thermal comfort reasoning using a lightweight LLaMA-3.2-3B student model. The framework consists of three sequential stages: (1) knowledge distillation from a teacher model, GPT-4o, to transfer domain-specific explanatory patterns; (2) supervised fine-tuning with ground-truth anchoring and a TSV-weighted loss to improve thermal sensation vote (TSV) prediction while limiting teacher-error propagation; and (3) verifier-guided direct preference optimization using an automated claim-consistency verifier. The verifier extracts factor-direction claims from generated explanations, such as high air temperature or low air speed, and checks them against comfort-centered percentile thresholds derived from the ASHRAE Global Thermal Comfort Database II. On the held-out test split, the final model achieved 42.7 % exact TSV accuracy and 79.9 % ±1 accuracy, comparable to standard tabular baselines on the same split. More importantly, the verifier-guided alignment stage reduced the any-incorrect-claim rate from 52.3 % to 16.5 % and improved linked claim accuracy to 78.3 %. Robustness tests showed reasonable directional consistency under perturbation without strong dependence on PMV or SET proxies. At the hot and cold tails, 81 % of explanations named an HVAC-controllable factor, so compact language models can pair thermal comfort predictions with verifiable, operator-actionable explanations for building operation.