L1 filename (sample_meta) → L3 audio fallback (sample_classify ensemble-fine, the 73% scoreboard winner) → L2 folder cross-check (flags folder_agrees=False when the folder name lies). Audio runs ONLY on opaque files. Per folder: dominant, agreement conf, homogeneous, kit_like, per-:index labels with source, provenance. Validates the domain knowledge mechanically: jazz=kit (000_BD→kick :0, HH/OH→hat, SN→snare), cpluck=multisample (plucks+body, folder='keys' flagged), 808hc=perc via audio.
