nx_gmmlu_research_fetch.nx

nx_gmmlu_research_fetch.nx -- F786 rung-1: ingest the PUBLIC Global-MMLU-Lite benchmark (a LITERAL row on the model card, b16) so the reasoning/chat battery is scored on REAL public data, never on questions I invented (self-authored questions = a self-quiz, not a benchmark -- scale law, no toys). Fetches the dataset contract + rows pages over own TLS-1.3 to knowledge/fetched/gmmlu_*.raw. Named *research_fetch* so nx_job_run's family gate can launch it ASYNC (pages exceed the 64KB MCP envelope and the whole run exceeds the 15s edge window). ff_putn writes its digits (seq206 class fixed). ⚠DATASET NAME: CohereForAI/Global-MMLU-Lite is RENAMED -> CohereLabs/Global-MMLU-Lite (the old name returns 404 RenamedDatasetError; verified live 2026-07-20). ⚠config=en is ASSUMED (Global-MMLU-Lite is per-language); the splits fetch above it exists precisely to CONFIRM the real config id -- read knowledge/fetched/gmmlu_splits.raw before trusting the rows page. expect_exit: 0 license_tier: ORIGINAL No hw writes (Rule 26).