According to BNPB data updated on August 29, 2026, the earthquake caused 111 deaths and displaced hundreds of thousands of people across seven regencies in NTT. ^[1] Master's students from the Sekolah Tinggi Ilmu Kepolisian (STIK) Lemdiklat Polri, 16th cohort, distributed social aid to earthquake victims in East Nusa Tenggara (NTT) on August 29, 2026. ^[2] The study demonstrates that behaviorally effective LoRA writes are sparse, structured, and far more concentrated than the raw low-rank parameterization suggests. ^[3] Agustinus Nong Viles, a camp coordinator in Dusun Wairhabi, received the aid on behalf of 159 households from three villages: Dusun Wairhabi, Dusun Habigahar, and Kelurahan Waiot. ^[4] The aid was delivered to displacement camps in Dusun Wairhabi, Desa Habi, Kecamatan Kangae, Kabupaten Sikka, and via Lorens Say port to Pelabuhan Palue. ^[5] Across GSM8K, MathQA, and AQuA benchmarks, per-module top-k continuation reached its optimum at k in {2, 4} in all twelve seed-level cases tested. ^[6] Researchers used Learned-Basis LoRA, a learned-basis continuation recipe, to expose the structure of LoRA writes. ^[7] The Learned-Basis LoRA recipe warms up an unconstrained adapter, converts its learned write columns into a module-wise orthonormal basis, freezes that basis, and continues training inside the constrained parameterization. ^[8]