Correct the cross-lingual figures: they came from a different checkpoint

README.md

@@ -46,7 +46,7 @@ training data, same recipe; the difference is the encoder underneath.
 | Test accuracy | **86.95%** | 86.72% |
 | Weighted F1 | **86.39%** | 86.33% |
 | Macro F1 | 83.06% | 82.95% |
-| Same sector for a README and its own English translation | **73%** | 19% |
+| Same sector for a README and its own English translation | **77%** | 19% |
 | Calibrated as trained | yes, `T` = 1.04 | yes, `T` = 1.01 |
 
 On English text the two are indistinguishable: a 0.2-point difference on a
@@ -63,22 +63,24 @@ brings is a shared multilingual representation space from its own pre-training,
 so a Spanish or Chinese README lands near its English equivalent and the
 classification head, trained on English, still applies.
 
-That transfer is real but imperfect, and it is what the 73% above measures: 93
+That transfer is real but imperfect, and it is what the 77% above measures: 93
 non-English repositories were classified twice, once as written and once from a
-human-quality English translation, and the two answers agreed 73% of the time.
+human-quality English translation, and the two answers agreed 77% of the time.
 RoBERTa-large agrees with itself 19% of the time on the same test, which is what
 having no vocabulary for the text looks like.
 
 | | FR (31) | ES (28) | RU (14) | PT (11) | ZH (9) | all (93) |
 |---|---:|---:|---:|---:|---:|---:|
-| this model | 84% | 57% | 79% | 91% | 56% | **73%** |
+| this model | 84% | 68% | 100% | 73% | 56% | **77%** |
 | RoBERTa-large | 16% | 11% | 21% | 18% | 56% | 19% |
 
-Two caveats worth stating plainly. **Agreement is not accuracy**: when the two
-readings differ, neither is known to be right. And the figure varies between
-training runs more than the English metrics do — a second run of the same recipe
-measured 81%, against an English F1 of 85.2. Treat 73% as one measurement, not a
-constant.
+Three caveats worth stating plainly. **Agreement is not accuracy**: when the two
+readings differ, neither is known to be right. **The per-language cells are tiny**
+— 9 to 31 repositories each, so a single flipped repository moves Chinese by 11
+points and the Russian 100% rests on 14 cases. Only the 93-repository total is
+worth quoting. And the figure varies between training runs more than the English
+metrics do — a second run of the same recipe measured 81%, against an English F1
+of 85.2. Treat 77% as one measurement, not a constant.
 
 Training on multilingual labelled data would be the real upgrade, and nothing
 here measures what it would add.
@@ -202,8 +204,9 @@ split or sequence length.
 ## Limitations
 
 - **The training data is English.** See the section above.
-- **Agreement across languages is not accuracy**, and it varies between runs
-  (73% here, 81% in a second run of the same recipe).
+- **Agreement across languages is not accuracy**, it rests on 93 repositories,
+  and it varies between runs (77% here, 81% in a second run of the same recipe).
+  The per-language breakdown has 9 to 31 repositories per cell.
 - **The labels are GPT-4.1 judgements**, not human annotation.
 - **The model cannot abstain.** The training data contains only positives, so
   every repository is assigned some sector. On a production corpus of public