Patients with depression exhibit a wide range of symptom profiles, yet clinical practice often collapses this variability into a single severity score. Large language models (LLMs) could, in principle, capture multiple symptoms and their intensities from patient speech, but how these depressive symptoms are represented inside the model remains unclear, limiting clinical trust.\ \ We applied mechanistic interpretability techniques to the residual stream of Gemma-3-27B-PT. By recording activations for symptom descriptions drawn from validated clinical instruments, we evaluated geometric separation across layers using several distance metrics and found the strongest separation at layer 21.\ \ Using Semantic Projection, we projected held‑out naturalistic text onto Symptom Vectors constructed from the same instruments. The resulting per‑symptom coefficients preserved the clinician‑annotated rank ordering across mood, somatic, and suicidality dimensions.\ \ Moreover, a single depression vector in layer 21 distinguished depressive from non‑depressive text (AUC = 0.789). This vector can serve as an emotional valence gate that restricts symptom projection to speech identified as depressive.\ \ These findings reveal a decorrelated, clinician‑aligned symptom signal that can be read directly from internal activations, providing a mechanistic foundation for interpretable depression‑assessment tools.\ \ Review