This paper investigates whether the domain of a LoRA fine‑tune can be recovered when only model artifacts are available—specifically weight deltas, internal activations on a fixed probe set, or generated text. Experiments use a single base model (Qwen3.5‑0.8B) to produce 36 LoRA fine‑tunes over 18 public datasets spanning five domains: medical, code, mathematics, non‑English text, and narrative prose. Three feature families are extracted at decreasing levels of access: white‑box weight statistics, gray‑box activations, and black‑box outputs. A pre‑registered leave‑one‑dataset‑out (LODO) protocol, frozen before any results, tests domain identification on previously unseen datasets. Results show domain recovery well above chance (20%): gray‑box activations achieve 61% accuracy, black‑box outputs 64%, and white‑box weight statistics 42%. Contrary to the hypothesis that signal decays monotonically from weights to activations to outputs, activations and outputs are more informative. The black‑box performance remains unchanged (64%) when only generated text is used, without token probabilities. Accuracy is stable across a third independent run per dataset, different data subsamples, and LoRA ranks ranging from 8 to 64. All code and per‑run feature vectors are released.\n\nReview: The study reveals that model artifacts leak far more domain information than previously assumed, raising new privacy concerns for fine‑tuned models and providing a solid benchmark for future mitigation strategies.