nx_vsbench.nx

nx_vsbench.nx -- THE HEAD-TO-HEAD VS-RULER: are we BETTER than the reference titles, or do we merely have the same feature LIST? nx_gamebench scores capability PRESENCE (HAVE/PARTIAL/GAP) against each title's declared vector. Presence can prove PARITY OF FEATURE LIST and nothing more -- its own honest block already says so: titles_reference_run = 0. This organ is the instrument that makes "better" ARITHMETIC instead of opinion. THE CONTRACT: a head-to-head CELL is (title, axis, ours_permil, theirs_permil, evidence_tier). tier 3 EXT = measured from a captured artifact that EXISTS ON DISK (the only tier that counts) tier 2 DOC = a published spec/manual figure tier 1 DECL = our own declaration (weakest -- this is what a self-graded board is made of) The headline beat_index_permil is computed over TIER-3 CELLS ONLY. Declarations can never raise it. With ZERO ext cells the headline is 0 and the verdict is UNGROUNDED -- which is exactly what it must say today, because we have never run a reference title head-to-head. Everything we BELIEVE about our structural advantages is emitted as HYPOTHESIS-NOT-CLAIM until a cell carries real evidence. That is the cynical-instruments law applied to the one number that could flatter us most. A SECOND QUEUE, DELIBERATELY: nx_gamebench's gap_queue ranks by titles_blocked = a PARITY objective ("what unblocks the most titles"). It structurally cannot answer "what makes us BETTER" -- which is why genetics-breeding sits near the bottom there (only Breeders needs it) while being one of the few axes where a generative emitter can beat a hand-authored roster outright. vsbench publishes the ADVANTAGE view beside it. Two objectives, two queues, both honest. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0