Model card: correct the replication section (run-order split caveat, authors' code vs port)

README.md

@@ -84,20 +84,30 @@ Use that loss to match the reference training.
 
 ## Replication (braindecode PR #1298)
 
-Fine-tuning on the 61-word task with the paper's protocol (200 epochs, AdamW,
-per-word 80/10/10 split, 6 seeds × 12 subjects, mean ± SE over subjects):
+Fine-tuning on the 61-word task with the authors' `run_cls.py` protocol (200
+epochs, AdamW, sigmoid cross-entropy, 6 seeds × 12 subjects, paper-time data
+revision `056a87b7`, mean ± SE over subjects). The authors' fine-tuning loop
+was run with both the original model and the braindecode port, starting from
+the same weights:
 
-| | paper (Table 2) | braindecode, sigmoid loss | braindecode, logits |
+| | paper (Table 2) | authors' code | braindecode port |
 |---|---|---|---|
-| Du-IN (mae) | 62.70 ± 4.69 | 54.51 ± 4.77 | 50.07 ± 5.26 |
-| Du-IN (vqvae) | 58.24 ± 4.83 | 49.33 ± 4.42 | 49.31 ± 5.04 |
-| scratch | 56.29 ± 5.20 | 48.40 ± 4.98 | 47.32 ± 5.26 |
-
-The pretraining gain over scratch is reproduced (+6.1 vs +6.4 in the paper,
-12/12 subjects), but absolute accuracies are about 8 points lower. The authors'
-own `run_cls.py` on the same released data gives the same numbers as the port
-(subject 002: 47.8 % mae vs 83.6 % in the paper), so the gap comes from the
-released data and checkpoints, not from the conversion.
+| Du-IN (mae) | 62.70 ± 4.69 | 60.52 ± 4.58 | 60.16 ± 4.62 |
+| Du-IN (vqvae) | 58.24 ± 4.83 | 55.10 ± 4.51 | 55.00 ± 4.51 |
+| scratch | 56.29 ± 5.20 | 54.06 ± 5.12 | 53.73 ± 5.13 |
+
+The port matches the authors' code to within 0.4 points. Training the port on
+plain logits (`cross_entropy(logits)`) instead of the sigmoid loss gives 55.6
+for mae.
+
+**Split caveat.** The authors' split concatenates a subject's runs in
+`os.listdir` order and puts the first 80 % of each word's trials in the
+training set. Validation and test therefore come almost entirely from the
+**last run listed**, and that run depends on the filesystem. If it is the
+chronologically last session (`run6`), accuracy drops sharply. On subject 002
+(scratch, seed 42), holding out `run6` gives 43–45 %, while holding out any
+other run gives 71–78 %. A sorted run order underestimates the paper by about
+8 points on average.
 
 ## License