SubmissionNumber#=%=#19 FinalPaperTitle#=%=#PerceptionLab at PsyDefDetect: Overcoming LLM Polarization via Rubric-Grounded Retrieval and Supervised Clinical Reasoning Distillation for Fine-Grained Ordinal Classification ShortPaperTitle#=%=# NumberOfPages#=%=# CopyrightSigned#=%=# JobTitle#==# Organization#==# Abstract#==#Automating the classification of psychological defense mechanisms is a critical yet challenging frontier in clinical natural language processing. General-purpose Large Language Models (LLM) struggle to apply fine-grained ordinal frameworks like the Defense Mechanism Rating Scales due to the implicit nature of clinical cues and a fundamental clinical reasoning gap that causes models to polarize predictions toward extreme labels. In this paper, we present our third-place system for the PsyDefDetect Shared Task at BioNLP 2026. Our best-performing system synergizes dynamic DMRS-Q items retrieval, performed by a larger Pro model, with a Gemini 2.5 Flash classifier fine-tuned via reasoning traces distilled from that same Pro model. This dual approach, grounding decisions in dynamic rubric criteria while leveraging task-specific supervision, successfully neutralizes prediction polarization. Our hybrid system achieves an accuracy of 67.37% and a macro-F1 of 39.56, providing empirical evidence that targeted clinical supervision paired with retrieval significantly outperforms the raw parameter scale of un-tuned foundation models. Author{1}{Firstname}#=%=#Tamjid Hasan Author{1}{Lastname}#=%=#Fahim Author{1}{Username}#=%=#tamjid Author{1}{Orcid}#=%=# Author{1}{Email}#=%=#tamjidhf@gmail.com Author{1}{Affiliation}#=%=#Rajshahi University of Engineering & Technology Author{2}{Firstname}#=%=#Syed Asif Author{2}{Lastname}#=%=#Johan Author{2}{Username}#=%=#q3alpha Author{2}{Orcid}#=%=# Author{2}{Email}#=%=#connect.syedasifjohan@gmail.com Author{2}{Affiliation}#=%=#Rajshahi University of Engineering and Technology Author{3}{Firstname}#=%=#Saad Author{3}{Lastname}#=%=#Bin Maksud Author{3}{Username}#=%=#saadov Author{3}{Orcid}#=%=# Author{3}{Email}#=%=#sbinmaksud@gmail.com Author{3}{Affiliation}#=%=#Rajshahi University of Engineering & Technology ========== èéáğö