# trio-r64 -- INFORMAL / NON-GATE

Offline CPU merge of 3 rank-64 LoRA adapters:
  0.3333333333333333  bach  adapters/bach-shakedown-r64/final
  0.3333333333333333  chopin  adapters/20260829T162506Z-stage0-r64/final
  0.3333333333333333  sousa  adapters/20260830T004943Z-sousa0-r64/final

Method: targets M = sum_i w_i * B_i @ A_i exactly, then takes the best
rank-64 representation of it by truncated SVD (equivalent to peft
add_weighted_adapter combination_type='svd', computed directly from the
adapter files so no base model is loaded). A per-key average of the LoRA
factors was rejected: it introduces cross terms and is not a blend.

Rank-64 truncation cost vs the exact rank-192 sum:
  mean 0.0150   min 0.0020   max 0.0430 (at base_model.model.layers.17.self_attn.k_proj)