nx_reader_squad_gate.nx

nx_reader_squad_gate.nx -- RUNG 3 (a/b/c) of the NEURAL PASSAGE READER arc: the reader trains on REAL SQuAD. R3c = ERROR-FEEDBACK SGD (nfa_sgd_ef): R3b's three runs all froze in a Q16 uniform-collapse absorbing state because sub-quantum updates truncate to zero; error-feedback keeps the residuals so no gradient is lost. R3a (2889ff525f) measured the from-scratch wall honestly: the pipeline + STABLE Q16 training work (grad-clip + LR-decay; loss 8401->2286) but at 664 rows the reader MEMORIZES without out-generalizing a question-blind token-prior baseline -- because 131k of 133k params were COLD-START embeddings. R3b (this version) = the measured lever: PRETRAINED SOVEREIGN EMBEDDINGS. The vocab is exact-word (61-bit ehash, same hash as the PPMI/SGNS vocab); the embedding table is INITIALIZED from R1's trained SGNS embeddings (knowledge/index/embed_v1.bin, full-data run) and FROZEN -- trainable params drop 133k -> ~7k (attention block + FFN + span heads only), attacking the memorization directly. Attention alignment (Wq/Wk bilinear) now operates on MEANINGFUL word geometry (question<->context similarity = SGNS cosine). T1 held-out span token-F1: reader > question-blind ablation + 100 permille (liar-kill at pretrained parity) T2 held-out exact-start > 60 permille (chance ~11) T3 pretraining helps: reader F1 > R3a's from-scratch 56 + 50 T4 deterministic re-eval Consumes: reader_rows.bin (nx_qabench 'dt'), semppmi_v1.bin (vocab hashes), embed_v1.bin (SGNS Q10). expect_exit: 0 license_tier: ORIGINAL Sovereign: nx_nofloat_autograd + nx_syscalls.