Sync model card: reproducibility scoping, seed-protocol precision

README.md

@@ -90,11 +90,14 @@ Both models sit at the ceiling of this benchmark — FLORES+ sentences are long
 1,012-way retrieval saturates. Read this as evidence of zero out-of-domain degradation, not as a
 margin over LaBSE.
 
-Training was repeated with three random seeds; Creole→English test ndcg@10 across seeds:
-**0.9653 ± 0.0002** (accuracy@1 **0.9433 ± 0.0006**). The released checkpoint is seed 42, designated
-before results were seen.
+The contrastive stage was repeated with three random seeds over the same deterministically mined
+negative set; Creole→English test ndcg@10 across seeds: **0.9653 ± 0.0002** (accuracy@1
+**0.9433 ± 0.0006**). The released checkpoint is seed 42, designated before results were seen.
 
-Every number is reproducible from the [training repository](https://github.com/LK-maker-007/morisien-embed).
+Every number in the tables above is reproducible from the
+[training repository](https://github.com/LK-maker-007/morisien-embed) (Matryoshka figures via
+`scripts/evaluate.py --truncate-dim`). The Haitian-proximity and case-sensitivity figures under
+Limitations come from an adversarial audit of the released checkpoint.
 
 ## Training
 
@@ -135,7 +138,10 @@ Every number is reproducible from the [training repository](https://github.com/L
   (cosine ≈ 0.81 to the same sentence); caps-heavy text retrieves worse.
 - **Protocol note.** During recipe development the held-out test score was printed at the end of each
   training run, so recipe selection had test visibility; an independent audit bounded the resulting
-  optimism at ≤ ~0.01 ndcg. The 3-seed replication was run after the recipe was frozen.
+  optimism at ≤ ~0.01 ndcg. The 3-seed replication was run after the recipe was frozen. Leak filtering
+  reserves the Creole side of every evaluation pair; English/French target texts are not reserved, and
+  an audit found 1 of 999 benchmark passages also occurring in training as the translation of a
+  different Creole sentence.
 - **Orthographic variation.** Training data mixes pre- and post-2011 (Lortograf Kreol Morisien)
   spellings; performance on older orthography is untested.