{ "checkpoint": "llama-3b_layer27_maghead_mlp_h2048x2_t0.03_a0.5.pt", "arch": "maghead_mlp (InfoNCE dir head [MLP 3072->2048->2048->1024, GELU] + MSE mag head)", "model": "meta-llama/Llama-3.2-3B-Instruct", "layer": 27, "data_shards": 24, "val_rank1": 0.631, "val_dir_rank1": 0.646, "val_cos": 0.377, "decode_cos_pred_sbert": 0.371, "decode_cos_upper_sbert": 0.944, "decode_cos_sonar_shortcut": 0.379, "n_decode_eval": 300, "beats": "linear maghead L27 24sh (rank1 0.594, decode 0.363) \u2014 MLP dir head, +6.2% rel rank1, decode held", "date": "2026-06-06", "note": "MLP direction head breaks the linear rank-1 ceiling without losing decode coherence." }