Upload README.md with huggingface_hub

README.md

@@ -58,6 +58,25 @@ Base and tuned generated in one process under identical greedy decoding, scored
 | enumerated calendar | 200 | 23.5% | **90.5%** |
 | range slice, 2031 to 2035 | 150 | 14.7% | **76.7%** |
 
+## Significance, stated rather than assumed
+
+Paired design, so significance is an **exact McNemar test** on the rows where exactly one
+model was right, with the null that the adapter is no better than the base.
+
+| slice | only tuned right | only base right | exact two-sided p |
+|---|---|---|---|
+| headline held-out | 405 | 5 | 7.2e-113 |
+| hard shard | 241 | 5 | 1.3e-64 |
+| enumerated calendar | 135 | 1 | 3.1e-39 |
+| range slice, 2031 to 2035 | 95 | 2 | 6.0e-26 |
+
+Note the middle column, which is **not** zero here. Alone in this portfolio, this base wins
+rows outright: 13 of them across the four slices. That is the same fact the top of this card
+opens with, arriving independently from the test rather than from the prose. A base that
+scores 69 to 86% on plain day-walking should win some rows, and it does.
+
+Recompute with the released `significance.py` from the released generation files.
+
 ## The gate that nearly killed this entry, and the rule written before it ran
 
 The pre-registered Day-0 gate measured a **blended** F1 floor of 34%, which sat between the pass and