nx_writebench_gate.nx

nx_writebench_gate.nx -- the LIAR-KILL gate for nx_writebench. A benchmark that grades us is worthless unless something can prove it is not a rubber stamp. Teeth: T1 THE RULER IS REAL -- the banked corpus exists and clears the decoy byte floor. T2 THE RULER IS THAT PAPER -- the corpus literally contains the paper's own distinctive strings (FlowGPT / Krippendorff / the H-/C+ finding). This is the anti-fabrication tooth: it makes it impossible to benchmark against an INVENTED paper, which is the whole failure mode of a self-authored ruler. T3 EVERY CLAIMED ARTIFACT EXISTS -- each evidence artifact the board attaches to a HAVE/PARTIAL is stat'd on disk from this gate's OWN independent list. Board, gate and disk are three copies. T4 NEG-CONTROL -- a fabricated artifact path must NOT exist. Without this, T3 could pass by accident (if the stat helper always returned success) and the gate would be a rubber stamp. T5 HONESTY CANNOT ROT -- the board source must still carry the SELF-GRADED deflation AND the DAN-class policy declaration. If a later edit quietly deletes either so the number reads like parity, this gate goes RED. T6 COVERAGE IS COMPUTED -- this gate recomputes the score from its own independent copy of the 28 verdicts and asserts it equals both the arithmetic (20/56 = 357 permil) AND the figure the board actually emitted into its log. Inflating the board alone cannot move this. T7 THE PARTITION SUMS -- the ruler's four class sizes must sum to 378, and 378 minus the 2 dual-labelled bots must equal the 376 distinct corpus. A partition is a claim: check the parts. T8 RULER DATA IS NON-VACUOUS -- the headline datum the board grades against (22.8) must be present in the banked corpus, not just in our prose about it. Evidence -> stdout + knowledge/status/writebench_gate.log. Exit 0 GREEN / 1 RED. Sovereign x86_64. license_tier: ORIGINAL expect_exit: 0