nx_bert_ce_lib.nx

nx_bert_ce_lib.nx -- SOVEREIGN BERT cross-encoder (BertForSequenceClassification with one logit) composed from the estate's own f32 tower: nx_safetensors_load (weights), vit_mhsa and vit_mhsa_pool (multi-head attention with biases, the ViT block's kernel), nx_f32_layernorm, nx_f32_matmul_t and its pooled twin, nx_f32_gelu_vec and nx_f32_tanh. The model is a LICENSED DATA ASSET (Apache-2.0, mirrored with provenance under knowledge/fetched/) read by NishiLang from the first byte; nothing foreign executes. BERT differs from the ViT block already in the tree in three places and only those are written here: POST-norm residuals (LayerNorm after each add), an embedding layer of word + absolute position + token type followed by LayerNorm, and the pooler (tanh of a dense over the [CLS] row) with the classifier. Depth, width, heads, intermediate size, position count and the LayerNorm epsilon are read from the model's config.json at load, never from a constant. Search rung R0, 2026-09-14. license_tier: ORIGINAL