nx_bright_reasoned_import.nx

nx_bright_reasoned_import.nx -- REFEREE-REPLAY IMPORTER for a published reasoned-query set (search R0l, 2026-09-16). A query reasoner's authors publish their rewrites as ONE JSON object keyed by the ORIGINAL BRIGHT query text with the rewrite as the value (TongSearch-QR ships one file per model, MIT). This organ matches that object, split by split, against the estate's own BEIR-shaped queries.tsv rows (written by nx_bright_prep, whose cells have every tab, newline and carriage return replaced by a space) and writes <root>/<split>/<outname> rows `id TAB text`: the alternate query file nx_beir_eval takes as `queries=<outname>`, which it scores as a REFEREE REPLAY and never writes to the leaderboard (a query another model wrote is not the estate's system). Two modes: `reason` (the rewrite alone is the query, the BRIGHT reasoning-query protocol the paper's Table 1 uses) and `concat` (the original query, one space, the rewrite). The JSON decoder flattens \n \t \r to a space so a decoded key compares BYTE-EQUAL to the prep organ's cell, encodes \uXXXX (surrogate pairs included) through the estate's utf8 encoder and passes raw UTF-8 through untouched. It REFUSES (exit 1, nothing renamed into place) when any split row has no rewrite, naming every missing id: a partial replay scores a different query set than the one it claims. Counts print as a partition (rows = matched + missed) beside the JSON pair count, and keys reused across splits are counted separately, because BRIGHT's two theoremqa splits share question texts, so a text-keyed set is SMALLER than the row count by construction and the two numbers must not be read as a mismatch. argv: <reasoned.json> <bright-root> <outname> [reason|concat] (the twelve splits are nx_beir_eval's) license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0 full match, 1 refused (named), 2 usage