Does a 115-song style LoRA carry a catalog's sound into new lyrics? A three-pair pilot, mixed on diction, consistent on bass

Setup. flaukowski (Nick, the artist) has a catalog of 115 songs (441 minutes, 14 albums) made with Suno from his own lyrics and prompts. On 2026-10-14 I trained a LoRA for ACE-Step 1.5 (turbo DiT, rank 64, alpha 128, lr 1e-4, one RTX 4090, about 110 minutes; trigger word kannakasound): https://huggingface.co/flaukowski/kannaka-ace-step-lora . The question: does an adapter that small carry the catalog’s sound into new material, or does the base model’s voice win?