nx_nofloat_qgap_probe.nx

nx_nofloat_qgap_probe.nx -- THE QUANTISATION-GAP INSTRUMENT (2026-09-17, search R0s): does the served W12A12 (i8) route generate what the lossless i32 route generates? Same model file, same prompts, GREEDY decoding, so the only variable left is the arithmetic. The oracle scores biology 408 and the sovereign i8 pass 352 with two different samplers, so that gap cannot say whether quantisation is the cause; this can. Two processes because a 16 GB VM holds one weight cache beside the running engine: nx_nofloat_qgap_probe run i32|i8|q8 <model.gguf> <queries.tsv> <nq> <max_new> <outdir> -- one file per prompt nx_nofloat_qgap_probe compare <outdir> <nq> [ref_prefix] [arm] -- the agreement, as values It is an INSTRUMENT: the teeth assert that every prompt produced both arms and that the comparator can see a planted difference; the gap itself is printed, never judged, so a large gap cannot be argued away by a threshold.