nx_reasoning.nx

nx_reasoning.nx -- sovereign TEST-TIME COMPUTE organ (self-consistency). The no.1 2025/26 frontier gap on the LLM census (nx_swcompare_gapmap llm, momentum=11): spend MORE inference compute per question to get a BETTER answer. This organ implements the foundational primitive: Self-consistency majority vote (Wang et al. 2023, ICLR): sample N diverse completions at temperature T, extract each final answer, return the mode. Accuracy rises with N on tasks where the error mode is "diverse wrong answers, consistent right answer". Generation goes through the CHATML TEMPLATE at the TOKEN level (<|im_start|> injected as ids -- the nx_f32_llm_chat_test.nx proven path). MEASURED 2026-07-09: Qwen2.5-Instruct on a RAW continuation prompt ("47 + 38 =") emits garbage even with correct digit-split tokenization (verified by nx_tok_probe); the instruct model is only in-distribution INSIDE its template. Composes the EXISTING sampler (Hinton-2015 temperature + Fan-2018 top-k) + forward_v4 -- ZERO new model machinery. The answers array is the composition point for future best-of-N verifier re-ranking. genealogy_id: wang_2023_self_consistency lineage_id: substrate_reasoning_v2 (v1 = raw-continuation, measured-hostile)