_t4_first_model_authored.nx

_t4_first_model_authored.nx -- T4 of the training-substrate ladder: the FIRST REAL MODEL. Composes the GATED T6/T5 core (nx_tgrad_core.nx, extracted verbatim from _tensor_grad_authored.nx) into two trained models, all bits-up sovereign f32, zero randomness: LANE 1 (nonconvex proof): XOR via a 2-4-1 relu MLP + AdamW, deterministic symmetry-breaking init (the zero-init relu trap documented at T6 is closed by a FIXED pattern table, not RNG -- gate-proven here; the principled-init research row stays open in the roadmap). LANE 2 (REAL TEAM DATA): 8-way service-class model on the ACTUAL elder census (knowledge/status/elder_census.log) -- first 6 name bytes -> {svc,gen,data,orch,worker,mon,util,other}. PARSE ORACLE: the census file carries its own ground truth (CENSUS-SUM per-class counts); the parsed+deduped dataset must match it EXACTLY before training is even attempted. GATES: O census-parse==CENSUS-SUM oracle | G XOR learned (loss<1/100, 4/4 preds within 1/4) | H census train accuracy == services (100%, per-class pred counts == ground truth) | I BIT-EXACT two-run training, BOTH lanes (the standing EXCEED axis). Evidence: loss curve + verdicts appended to knowledge/status/modelwright.log (MODELWRIGHT rows = the role's first durable evidence; the scorecard reads this file). LAWS: struct-free, flat ifs, no &&/||, no unary-minus literals, no randomness. license_tier: ORIGINAL