The embedding store now carries 134 keys

README.md

@@ -108,9 +108,8 @@ Two limits worth knowing before feeding it anything:
 - **Every training and test epitope is a 15-mer.** The arms were normalised to 15 residues. Other
   lengths run, since the encoder pads and masks, but they are outside the training distribution.
 - **Coverage is 116 chains**: 98 human HLA class II chains and 18 mouse H2 chains, the ones the four
-  arms actually use. The research embedding store carries 154 keys; the other 38 (20 BoLA, Mamu and
-  SLA chains, plus 18 further human alleles) were curated alongside them, and no released model was
-  trained to score them.
+  arms actually use. The research embedding store carries 134 keys; the other 18 are further human
+  alleles curated alongside them, and no released model was trained to score them.
 
 ## Usage
 
@@ -165,7 +164,7 @@ head unchanged.
 - **The embedding store moved after training.** `emb_hla_esmc_small_0430.h5` as published in
   [`daylight-00/prepibind-embeddings`](https://huggingface.co/datasets/daylight-00/prepibind-embeddings)
   received the 2026-09-08 H2 chain-swap correction and differs from what these checkpoints were
-  trained on in 3 of its 154 keys: `H2-IAdA` and `H2-IAdB`, which held each other's arrays, and
+  trained on in 3 of its 134 keys: `H2-IAdA` and `H2-IAdB`, which held each other's arrays, and
   `H2-IAg7A`, same shape and different values. The correction is disclosed, not retro-applied. All
   three are mouse H2 chains and all three are in the training data — 0.48 % to 0.66 % of the rows
   of each arm — so **predictions for H2-IAd and H2-IAg7 come from a model fit on the pre-fix