nx_vsbest_lib.nx

nx_vsbest_lib.nx -- THE COMPARATIVE AXIS: admit or refuse a vs-best-in-class claim. WHY THIS DID NOT EXIST AND WHY THAT MATTERED. nx_maturity_auditor defines level 4 S-CLASS as TRIANGULATED PARITY vs a named best-in-class and level 5 EXCEED as a triangulated WIN. Those are CLAIMS ABOUT A COMPARISON -- and there was nowhere to put one. The only thing that could reach those rungs was permil_to_level awarding them from a COVERAGE PERCENTAGE (capped at PRODUCTION on 2026-07-31). So the board reporting "S-class reached in 0 domains" was never evidence that nothing is good; it is evidence that THE CLAIM COULD NOT BE MADE HONESTLY. This lib is the missing place. A claim is ADMISSIBLE only if ALL FOUR hold: competitor -- a NAMED external artifact (gcc, llvm, qemu, seL4, Vortex 3.0), never "the baseline" method -- how it was run, stated well enough to be RE-RUN by someone else evidence -- a READABLE log in which the declared pattern ACTUALLY APPEARS verdict -- behind | parity | ahead ★★★★★A CLAIM MISSING ANY OF THESE IS REFUSED, NOT DOWNGRADED. An unverifiable comparison is not a weaker comparison, it is the ABSENCE of one -- the same DANGLING-vs-measured split em_derive_level enforces for capability evidence, applied to the comparative axis. ★★★★★AND THE COMPARATOR MUST BE NON-EMPTY BY NAME. The corpus already banks the failure this prevents: a census scored "vs 0 SOTA-src" -- an EMPTY comparator, self-referential, able only to drift upward. Comparing against nothing is how 1000-permil coverage coexists with PROVEN 0/40. license_tier: ORIGINAL No hw writes (Rule 26).