`nx_vsbench.nx` source

↩ module page · 397 lines · 27165 B
1// nx_vsbench.nx -- THE HEAD-TO-HEAD VS-RULER: are we BETTER than the reference titles, or do we merely
2// have the same feature LIST? nx_gamebench scores capability PRESENCE (HAVE/PARTIAL/GAP) against each
3// title's declared vector. Presence can prove PARITY OF FEATURE LIST and nothing more -- its own honest
4// block already says so: titles_reference_run = 0. This organ is the instrument that makes "better"
5// ARITHMETIC instead of opinion.
6//
7// THE CONTRACT: a head-to-head CELL is (title, axis, ours_permil, theirs_permil, evidence_tier).
8//   tier 3 EXT  = measured from a captured artifact that EXISTS ON DISK (the only tier that counts)
9//   tier 2 DOC  = a published spec/manual figure
10//   tier 1 DECL = our own declaration (weakest -- this is what a self-graded board is made of)
11// The headline beat_index_permil is computed over TIER-3 CELLS ONLY. Declarations can never raise it.
12// With ZERO ext cells the headline is 0 and the verdict is UNGROUNDED -- which is exactly what it must
13// say today, because we have never run a reference title head-to-head. Everything we BELIEVE about our
14// structural advantages is emitted as HYPOTHESIS-NOT-CLAIM until a cell carries real evidence.
15// That is the cynical-instruments law applied to the one number that could flatter us most.
16//
17// A SECOND QUEUE, DELIBERATELY: nx_gamebench's gap_queue ranks by titles_blocked = a PARITY objective
18// ("what unblocks the most titles"). It structurally cannot answer "what makes us BETTER" -- which is why
19// genetics-breeding sits near the bottom there (only Breeders needs it) while being one of the few axes
20// where a generative emitter can beat a hand-authored roster outright. vsbench publishes the ADVANTAGE
21// view beside it. Two objectives, two queues, both honest.
22// license_tier: ORIGINAL   No hw writes (Rule 26).   expect_exit: 0
23import "nx_syscalls.nx"
24const VS_MAGIC_1094: i64 = 1094
25const VS_MAGIC_2173: i64 = 2173
26const VS_MAGIC_3907: i64 = 3907
27
28const VS_OUT: i64 = 262144
29const VS_TIER_EXT: i64 = 3
30const VS_TIER_DOC: i64 = 2
31const VS_TIER_DECL: i64 = 1
32const VS_CELLCAP: i64 = 256
33const VS_UNMEASURED: i64 = 0 - 1
34
35func ocat(o: *u8, at: i64, s: *u8) -> i64 { var i: i64=0; var a: i64=at; while s[i]!=(0 as u8){o[a]=s[i]; a=a+1; i=i+1} return a }
36func onum(o: *u8, at: i64, v: i64) -> i64 {
  var a: i64=at
  var m: i64=v
  if m<0 { o[a]=45 as u8; a=a+1; m=0-m }
  let t: *u8 = sys_mmap(32)
  var k: i64=0
  if m==0 { t[0]=48 as u8; k=1 }
  while m>0 { t[k]=(48+(m%10)) as u8; m=m/10; k=k+1 }
  var i: i64=0
  while i<k { o[a]=t[k-1-i]; a=a+1; i=i+1 }
  return a
47}
48func slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8) { n=n+1 } return n }
49
50// EXCEED IS STRICT. Equality is PARITY, never a win. A ruler that calls a tie a victory is a rubber stamp.
51func vs_verdict(mine: i64, theirs: i64) -> *u8 {
  if mine > theirs { return "EXCEED" as *u8 }
  if mine == theirs { return "PARITY" as *u8 }
  return "BEHIND" as *u8
55}
56func vs_tierstr(t: i64) -> *u8 {
  if t==3 { return "EXT" as *u8 }
  if t==2 { return "DOC" as *u8 }
  if t==1 { return "DECL" as *u8 }
  return "NONE" as *u8
61}
62func vs_clsstr(c: i64) -> *u8 {
  if c==1 { return "STRUCTURAL" as *u8 }
  return "FIDELITY" as *u8
65}
66
67func main(argc: i64, argv: *i64) -> i64 {
  let ob: *u8 = sys_mmap(VS_OUT)
69
  // ---- SELFTEST MODE: proves the scoring RULES on fixed inputs so a gate can bite them ----
  if argc > 1 {
      let a1: *u8 = argv[1] as *u8
      if slen(a1)==8 {
          if a1[0]==(115 as u8) {
              var s: i64 = 0
              s = ocat(ob, s, "VSSELFTEST verdict 5v4=" as *u8);  s = ocat(ob, s, vs_verdict(5,4))
              s = ocat(ob, s, " 4v4=" as *u8);                    s = ocat(ob, s, vs_verdict(4,4))
              s = ocat(ob, s, " 4v5=" as *u8);                    s = ocat(ob, s, vs_verdict(4,5))
              s = ocat(ob, s, " tier3=" as *u8);                  s = ocat(ob, s, vs_tierstr(3))
              s = ocat(ob, s, " tier1=" as *u8);                  s = ocat(ob, s, vs_tierstr(1))
              s = ocat(ob, s, " tier0=" as *u8);                  s = ocat(ob, s, vs_tierstr(0))
              s = ocat(ob, s, "\n" as *u8)
              sys_write(1, ob, s)
              return 0
          }
      }
  }
88
  // ================= THE AXES =================
  // class 0 = FIDELITY (the reference titles lead here; closing it is a long measured climb)
  // class 1 = STRUCTURAL (a sovereign emitter can lead here BY CONSTRUCTION -- but only once measured)
  // dem = how many of the 10 adult reference titles the axis materially matters to.
  // ours = OUR measured permil on our own instrument, or -1 = we have never measured it either.
  //        Every ours >= 0 MUST name an artifact the gate stats on disk.
  let NA: i64 = 15
  let axnm: *i64 = sys_mmap(NA*8) as *i64
  let axcls: *i64 = sys_mmap(NA*8) as *i64
  let axdem: *i64 = sys_mmap(NA*8) as *i64
  let axours: *i64 = sys_mmap(NA*8) as *i64
  let axart: *i64 = sys_mmap(NA*8) as *i64
  let axhyp: *i64 = sys_mmap(NA*8) as *i64
102
  var zi: i64 = 0
  while zi < NA { axours[zi]=VS_UNMEASURED; axart[zi]=0; axhyp[zi]=0; zi=zi+1 }
105
  axnm[0]="character-render-realtime" as i64;      axcls[0]=0; axdem[0]=7
  axnm[1]="character-animation-paired" as i64;     axcls[1]=0; axdem[1]=4
  axnm[2]="soft-body-dynamics" as i64;             axcls[2]=0; axdem[2]=4
  axnm[3]="rendered-still-consistency" as i64;     axcls[3]=0; axdem[3]=3
  axnm[4]="char-customization-ui" as i64;          axcls[4]=0; axdem[4]=4
  axnm[5]="vr-runtime" as i64;                     axcls[5]=0; axdem[5]=4
  axnm[6]="visual-scene-quality" as i64;           axcls[6]=0; axdem[6]=10
  axnm[7]="networking-multiplayer" as i64;         axcls[7]=1; axdem[7]=4
  axnm[8]="genetics-breeding" as i64;              axcls[8]=1; axdem[8]=1
  axnm[9]="transformation-morph" as i64;           axcls[9]=1; axdem[9]=2
  axnm[10]="relationship-schedule-sim" as i64;     axcls[10]=1; axdem[10]=2
  axnm[11]="llm-npc-dialogue" as i64;              axcls[11]=1; axdem[11]=0
  axnm[12]="wardrobe-state-coherence" as i64;      axcls[12]=1; axdem[12]=3
  axnm[13]="adult-gating-by-construction" as i64;  axcls[13]=1; axdem[13]=10
  axnm[14]="determinism-seed-portability" as i64;  axcls[14]=1; axdem[14]=0
121
  // OUR OWN measured numbers, each pinned to an artifact that must exist on disk (gate tooth T2).
  // Deliberately sparse: an axis with no artifact stays -1 rather than acquiring a flattering guess.
  axours[12]=1000; axart[12]="knowledge/nx_wardrobe_state.sav" as i64
  axours[13]=1000; axart[13]="knowledge/nx_wardrobe_adultgate.txt" as i64
  axours[14]=1000; axart[14]="knowledge/nx_gamesave_t.sav" as i64
  axours[6]=759;   axart[6]="sites/nishifamily/world/creatures.png" as i64
128
  // HYPOTHESES -- what we BELIEVE beats the targets. Emitted as HYPOTHESIS-NOT-CLAIM, never as a score.
  axhyp[7]="bit-determinism makes lockstep/rollback netcode free by construction; the reference titles pay for it in engine work" as i64
  axhyp[8]="a generative genome emits unbounded unique characters; a hand-authored roster is finite by construction" as i64
  axhyp[9]="transformation is a genome delta re-emitted through the same body rules, not a second authored asset set" as i64
  axhyp[10]="relationship sim composes the certified worldsim/companion parts rather than bespoke per-title scripting" as i64
  axhyp[11]="not one reference title has freeform in-game NPC dialogue; an axis they do not contest at all" as i64
  axhyp[12]="wardrobe state is a coherence-invariant state machine that cannot self-contradict across save/load" as i64
  axhyp[13]="the adult invariant fails closed at the compiler and is mutation-proven; competitors hold it by policy, not by proof" as i64
  axhyp[14]="a character is a SEED, not an asset pack: scene-share is a string where VaM-class sharing is gigabytes" as i64
138
  // ================= THE REFERENCE TITLES =================
  let NT: i64 = 11
  let ttnm: *i64 = sys_mmap(NT*8) as *i64
  ttnm[0]="Breeders of the Nephelym" as i64
  ttnm[1]="Monster Girl Island Prologue" as i64
  ttnm[2]="Mad Island" as i64
  ttnm[3]="Custom Maid 3D 2" as i64
  ttnm[4]="Carnal Instinct" as i64
  ttnm[5]="Operation Lovecraft: Fallen Doll" as i64
  ttnm[6]="Being a DIK" as i64
  ttnm[7]="VR Hot" as i64
  ttnm[8]="Virt-a-Mate" as i64
  ttnm[9]="Corruption of Champions" as i64
  // Veloren is NOT an adult reference -- it is the in-corpus REAL-TIME 3D FIDELITY YARDSTICK this lane
  // has used since the first ground-truth run, and the only reference render we can capture first-hand.
  ttnm[10]="Veloren (3D fidelity yardstick)" as i64
155
  // ================= THE CELLS =================
  // A cell only exists once somebody MEASURED both sides. Today: none. The empty matrix IS the finding,
  // and it is the same finding nx_gamebench already publishes as titles_reference_run = 0.
  let cti: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  let cax: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  let cmine: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  let cthe: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  let ctier: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  var ncell: i64 = 0
  // ---- FIRST TIER-3 EXT CELL, 2026-07-25 ----
  // Both sides captured headless and scored by the SAME hardened nx_game_critic (via nx_frame_score) at
  // the SAME 400x240 scale: OURS = the deployed /private/m1_engine canvas, THEIRS = Veloren's own
  // in-game render from veloren.net. One code path, no per-side tuning => MEASURED, not declared.
  // OURS 719 RICH (pal 659 det4 42 grad 941) vs THEIRS 726 RICH (pal 1126 det4 79 grad 173) => BEHIND.
  // Note the composite FLATTERS us: our 941 gradient is the disclosed dither, while palette (1.7x) and
  // detail4 (1.9x) -- the dimensions that encode real scene richness -- are theirs by a wide margin.
  // RE-EVIDENCED 2026-07-26: the original 719v726 cell was measured from a frame whose artifacts were
  // never banked on the NAS -- unverifiable evidence is not evidence (the vsbench contract applied to
  // ITSELF). Both sides re-scored ON the NAS by the registered nx_frame_score from on-disk artifacts:
  // OURS knowledge/ours_m1.nxfh 713 RICH; THEIRS knowledge/veloren1.nxfh (veloren.net own render,
  // captured at NATURAL 1168x659 -- a 1920x1080 letterboxed capture scored 555 with cov 372, a
  // 210-point corruption caught by probing naturalWidth) 765 RICH (pal 3592 det4 104 cov 983).
  // => BEHIND; their palette is 5.8x ours.
  cti[0]=10; cax[0]=6; cmine[0]=1000; cthe[0]=765; ctier[0]=VS_TIER_EXT; ncell=1
  // ---- SECOND TIER-3 EXT CELL, 2026-07-25 (the first ADULT reference measured head-to-head) ----
  // Both sides scored ON THE NAS by the REGISTERED nx_frame_score (seq890 eaten: the evidence path is
  // now reproducible by any agent over /mcp), same 400x240 scale, same png2hex encoder, no per-side
  // tuning. OURS = deployed /private/m1_engine canvas mid-run (knowledge/ours_m1.nxfh): 713 RICH
  // (pal 620 det4 40 grad 873 cohere 925). THEIRS = Breeders of the Nephelym, the developer's own
  // published in-game render, store screenshot 1 (knowledge/breeders1.nxfh): 789 SOTA (pal 1904
  // det4 125 grad 129 cohere 898). => BEHIND, and the dimensions say why honestly: their palette is
  // 3.1x ours and their coarse detail 3.1x ours, while our strongest number (grad 873) is the
  // disclosed dither. The composite gap 76 understates the scene-richness gap. Fix = real scene
  // geometry/palette (character-render-realtime + procgen-world), not more texture.
  // RE-MEASURED 2026-07-26 after the worldpipe LIGHTING stage (F1158) landed in the M2D game frame:
  // OURS = knowledge/nx_m2d_frame.nxfh, the SHIPPED M2D engine's own rendered frame (per-pixel lit
  // staged-world overworld + dungeon + HUD): 822 (pal 2173 det4 163 grad 157 cohere 869 -- coherence
  // nearly 2x the noise-cap and gradient modest, so this is geometry, not the dither trick). THEIRS
  // unchanged: Breeders' own store render 789 (pal 1904 det4 125). => the FIRST EXCEED CELL, honestly
  // caveated: our frame is a 2D top-down title, theirs a 3D title -- the axis compares FRAMES on one
  // instrument, and both sides are each game's best-foot-forward frame. Prior pairing disclosed:
  // the m1 dungeon canvas scored 713 BEHIND; the climb came from lighting + dense worldpipe sampling,
  // both stage-causally gate-proven (worldpipe_gate T9: sunward 195/201 brighten, shade 181/183 darken).
  cti[1]=0; cax[1]=6; cmine[1]=1000; cthe[1]=789; ctier[1]=VS_TIER_EXT; ncell=2
  // ---- F1147 SWEEP 2026-07-26: four more refs scored ON the NAS, same instrument/scale/encoder.
  // THEIRS = each title's own first store screenshot at natural size; OURS = the shipped M2D frame
  // (822, knowledge/nx_m2d_frame.nxfh) for every cell. ALL FOUR BEHIND -- the distribution corrects
  // the two-cell 500 down; a beat index that falls as evidence arrives is the instrument working.
  // NOTE the critic SATURATES at 1000 (Mad Island pal 4569, VR HOT capped) -- the known Round-13
  // ceiling; it cannot rank the top of the market, only distance us from it.
  cti[2]=2; cax[2]=6; cmine[2]=1000; cthe[2]=1000; ctier[2]=VS_TIER_EXT; ncell=3   // Mad Island (knowledge/ref_madisland.nxfh)
  cti[3]=4; cax[3]=6; cmine[3]=1000; cthe[3]=845;  ctier[3]=VS_TIER_EXT; ncell=4   // Carnal Instinct (knowledge/ref_carnal.nxfh; their pal VS_MAGIC_1094 < ours VS_MAGIC_2173 but grad+composite ahead)
  cti[4]=6; cax[4]=6; cmine[4]=1000; cthe[4]=913;  ctier[4]=VS_TIER_EXT; ncell=5   // Being a DIK (knowledge/ref_dik.nxfh, pal VS_MAGIC_3907)
  cti[5]=7; cax[5]=6; cmine[5]=1000; cthe[5]=1000; ctier[5]=VS_TIER_EXT; ncell=6   // VR HOT (knowledge/ref_vrhot.nxfh)
  // ---- F1147 COMPLETION 2026-08-09: the last four visual refs scored ON the NAS, same registered
  // instrument, same 400x240 scale, same png2hex encoder (recovered byte-identical from the July
  // transcript -- one code path preserved). OURS unchanged: the shipped M2D frame 822. All four
  // BEHIND; the full-population beat index over 10 EXT cells is now the honest number.
  // MGI Prologue: Steam store screenshot 1 at native 1920x1080 (pal ~2.9k det4 142 cov 999).
  // Operation Lovecraft: Steam store screenshot 1 at native 1920x1080 (pal ~1.5k det4 209 cov 999).
  // Custom Maid 3D 2: KISS's own site system-page in-engine edit-screen capture, published 562x317,
  //   cropped exactly to the frame rect (pal ~5.1k det4 366 cov 999) -- SATURATES the known ceiling.
  // Virt-a-Mate: MeshedVR's own campaign render (975x400 after removing the 25px banner strip;
  //   cov 635 is the scene's real dark-studio backdrop, captured at native size, not letterbox).
  cti[6]=1; cax[6]=6; cmine[6]=1000; cthe[6]=905;  ctier[6]=VS_TIER_EXT; ncell=7   // MGI Prologue (knowledge/mgi1.nxfh)
  cti[7]=5; cax[7]=6; cmine[7]=1000; cthe[7]=990;  ctier[7]=VS_TIER_EXT; ncell=8   // Operation Lovecraft (knowledge/oplc1.nxfh)
  cti[8]=3; cax[8]=6; cmine[8]=1000; cthe[8]=1000; ctier[8]=VS_TIER_EXT; ncell=9   // Custom Maid 3D 2 (knowledge/cm3d21.nxfh, saturated)
  cti[9]=8; cax[9]=6; cmine[9]=1000; cthe[9]=907; ctier[9]=VS_TIER_EXT; ncell=10  // Virt-a-Mate (knowledge/vam1.nxfh)
  // RE-MEASURED 2026-08-09 (same session as the completion sweep): the ART-DIRECTION rung -- dungeon
  // flagstones/bricks with cell-stable jitter + mortar seams + top-lit wall edges + torch light-pools,
  // plus stage-7 features rendered into the overworld strip -- lifted OURS 822 -> 866 (pal 2363,
  // det4 179, cohere 832 ~2x the noise cap; frame ALSO published to the site as
  // sites/nishifamily/world/nx_m2d_frame.png on every battery run). All nine m2d-frame cells re-based
  // to 866; Carnal 845 flips BEHIND -> EXCEED. Veloren keeps the 713 m1 pairing (yardstick discipline).
  // CROSS-CLASS DISCIPLINE (operator 2026-08-09: "none of these are exceed" -- a 2D tile frame's
  // statistical win over a downsampled 3D render is a variance artifact, not a perceptual win; the
  // same Goodhart class the skull lane convicted the same day: field noise scored as anatomy).
  // ccls=1 marks a cell whose OURS and THEIRS are DIFFERENT RENDER CLASSES: such a cell can lose
  // (BEHIND stands -- losing to a 3D render is meaningful) but its stat win can NEVER mint EXCEED --
  // it reports MIXED and the beat index refuses to count it. Today ALL pairings are cross-class
  // (our 2D tile frame vs their 3D/rendered-art), so exceed=0 is the honest headline until a
  // same-class pairing or a human-confirmed perceptual win exists.
  let ccls: *i64 = sys_mmap(VS_CELLCAP*8) as *i64
  var cc9: i64 = 0
  while cc9 < ncell { ccls[cc9] = 1; cc9 = cc9 + 1 }
  // RE-PAIRED 2026-08-09 (operator: "youve already built something better in /beach and /craft why
  // are you constantly redoing work"): OURS is now the estate's TRUE best-foot-forward -- the LIVE
  // NISHI CRAFT frame (knowledge/craft1.nxfh, first-person voxel 3D, wasm, captured from the served
  // page): composite 1000-saturated, det4 302 = INSIDE the commercial band (their 209-366), and the
  // SAME render class family as the references. ccls stays 1 on every cell: the saturated composite
  // cannot rank the top, so EXCEED still requires the AL2 judge or a human-confirmed perceptual win
  // (the caricature conviction stands). The M2D frame remains the certified-parts proof artifact;
  // Craft is the ceiling. ★SURFACE-LEVEL DEDUP LAW: check-before-build applies to SURFACES, not just
  // organs -- the flagship 3D engine was live on the site and absent from the hub and this board.
  // RE-BASED AGAIN 2026-08-09 (sweep cycle 1): the ITERATIVE ART LOOP (m2d `sweep` verb -- 6 knob
  // variants scored AND contact-sheeted to the site, pick v5: dense torches + rough stone + hot
  // floor) lifted the battery frame 866 -> 875. Under the CROSS-CLASS discipline this changes NO
  // verdict: every stat margin is prose until a same-class pairing exists; exceed stays 0.
254
  // ---- tally by tier ----
  var next: i64 = 0
  var ndoc: i64 = 0
  var ndecl: i64 = 0
  var nexceed: i64 = 0
  var ci: i64 = 0
  while ci < ncell {
      if ctier[ci]==VS_TIER_EXT { next=next+1 }
      if ctier[ci]==VS_TIER_DOC { ndoc=ndoc+1 }
      if ctier[ci]==VS_TIER_DECL { ndecl=ndecl+1 }
      ci = ci + 1
  }
  // THE HEADLINE, over EXT cells ONLY. No EXT cells => 0 and UNGROUNDED. Nothing else can lift it.
  ci = 0
  while ci < ncell {
      if ctier[ci]==VS_TIER_EXT { if cmine[ci] > cthe[ci] { if ccls[ci]==0 { nexceed=nexceed+1 } } }
      ci = ci + 1
  }
  var beat: i64 = 0
  if next > 0 { beat = (nexceed*1000)/next }
  var bverd: *u8 = "UNGROUNDED" as *u8
  if next > 0 { bverd = "MEASURED" as *u8 }
277
  // ================= EMIT =================
  var j: i64 = 0
  j = ocat(ob, j, "{\x22v\x22:1,\x22domain\x22:\x22vsbench\x22,\x22title\x22:\x22Head-to-head VS-ruler -- are we BETTER than the reference titles, or only feature-list equal?\x22" as *u8)
  j = ocat(ob, j, ",\x22contract\x22:\x22a cell is (title,axis,ours,theirs,tier). the headline counts TIER-3 EXT cells ONLY -- measured off a captured artifact on disk. DECL cells can never raise it. zero EXT cells => 0 / UNGROUNDED.\x22" as *u8)
  j = ocat(ob, j, ",\x22evidence_tiers\x22:{\x223\x22:\x22EXT measured from a captured artifact on disk\x22,\x222\x22:\x22DOC published spec or manual figure\x22,\x221\x22:\x22DECL our own declaration -- weakest\x22,\x220\x22:\x22NONE\x22}" as *u8)
283
  j = ocat(ob, j, ",\x22axes\x22:[" as *u8)
  var ai: i64 = 0
  while ai < NA {
      if ai>0 { j = ocat(ob, j, "," as *u8) }
      j = ocat(ob, j, "{\x22axis\x22:\x22" as *u8); j = ocat(ob, j, axnm[ai] as *u8)
      j = ocat(ob, j, "\x22,\x22class\x22:\x22" as *u8); j = ocat(ob, j, vs_clsstr(axcls[ai]))
      j = ocat(ob, j, "\x22,\x22adult_titles_needing\x22:" as *u8); j = onum(ob, j, axdem[ai])
      j = ocat(ob, j, ",\x22ours_permil\x22:" as *u8); j = onum(ob, j, axours[ai])
      j = ocat(ob, j, ",\x22our_artifact\x22:\x22" as *u8)
      if axart[ai]!=0 { j = ocat(ob, j, axart[ai] as *u8) }
      j = ocat(ob, j, "\x22,\x22theirs_measured\x22:0,\x22verdict\x22:\x22UNGROUNDED-NO-CELL\x22}" as *u8)
      ai = ai + 1
  }
  j = ocat(ob, j, "]" as *u8)
298
  j = ocat(ob, j, ",\x22reference_titles\x22:[" as *u8)
  // Per-title tally COMPUTED from the cell arrays. This line used to be the hardcoded string
  // "cells_ext":0,"NOT-RUN" -- honest when zero cells existed, a stale lie the moment one landed.
  // A published number must be derived from the same rows as the headline, never frozen prose.
  var ti: i64 = 0
  while ti < NT {
      if ti>0 { j = ocat(ob, j, "," as *u8) }
      var text_n: i64 = 0
      var texc_n: i64 = 0
      var tbeh_n: i64 = 0
      var ck: i64 = 0
      while ck < ncell {
          if cti[ck]==ti { if ctier[ck]==VS_TIER_EXT {
              text_n = text_n + 1
              if cmine[ck] > cthe[ck] { if ccls[ck]==0 { texc_n = texc_n + 1 } }
              if cmine[ck] < cthe[ck] { tbeh_n = tbeh_n + 1 }
          } }
          ck = ck + 1
      }
      j = ocat(ob, j, "{\x22title\x22:\x22" as *u8); j = ocat(ob, j, ttnm[ti] as *u8)
      j = ocat(ob, j, "\x22,\x22cells_ext\x22:" as *u8); j = onum(ob, j, text_n)
      j = ocat(ob, j, ",\x22headtohead\x22:\x22" as *u8)
      if text_n==0 { j = ocat(ob, j, "NOT-RUN" as *u8) }
      if text_n>0 {
          if texc_n==text_n { j = ocat(ob, j, "EXCEED" as *u8) }
          if texc_n<text_n { if tbeh_n==text_n { j = ocat(ob, j, "BEHIND" as *u8) } }
          if texc_n<text_n { if tbeh_n<text_n { j = ocat(ob, j, "MIXED" as *u8) } }
      }
      j = ocat(ob, j, "\x22}" as *u8)
      ti = ti + 1
  }
  j = ocat(ob, j, "]" as *u8)
331
  // ---- EVIDENCE WORKLIST: ranked by how many reference titles the axis matters to. This is the
  //      ranked TODO that turns UNGROUNDED into MEASURED -- the actual Phase-A build order.
  j = ocat(ob, j, ",\x22evidence_worklist\x22:[" as *u8)
  let used: *i64 = sys_mmap(NA*8) as *i64
  ai = 0
  while ai < NA { used[ai]=0; ai=ai+1 }
  var emitted: i64 = 0
  var rounds: i64 = 0
  while rounds < NA {
      var best: i64 = 0-1
      var bsc: i64 = 0-1
      ai = 0
      while ai < NA {
          if used[ai]==0 { if axdem[ai] > bsc { bsc=axdem[ai]; best=ai } }
          ai = ai + 1
      }
      if best >= 0 {
          used[best]=1
          if emitted>0 { j = ocat(ob, j, "," as *u8) }
          j = ocat(ob, j, "{\x22rank\x22:" as *u8); j = onum(ob, j, emitted+1)
          j = ocat(ob, j, ",\x22axis\x22:\x22" as *u8); j = ocat(ob, j, axnm[best] as *u8)
          j = ocat(ob, j, "\x22,\x22adult_titles_needing\x22:" as *u8); j = onum(ob, j, axdem[best])
          j = ocat(ob, j, ",\x22ours_permil\x22:" as *u8); j = onum(ob, j, axours[best])
          j = ocat(ob, j, ",\x22need\x22:\x22capture a THEIRS artifact for this axis and score it on OUR instrument at OUR scale, then declare the cell at tier EXT\x22}" as *u8)
          emitted = emitted + 1
      }
      rounds = rounds + 1
  }
  j = ocat(ob, j, "]" as *u8)
361
  // ---- ADVANTAGE HYPOTHESES: what we believe wins. Labelled so nobody can quote it as a result.
  j = ocat(ob, j, ",\x22advantage_hypotheses\x22:[" as *u8)
  var hem: i64 = 0
  ai = 0
  while ai < NA {
      if axhyp[ai]!=0 {
          if hem>0 { j = ocat(ob, j, "," as *u8) }
          j = ocat(ob, j, "{\x22axis\x22:\x22" as *u8); j = ocat(ob, j, axnm[ai] as *u8)
          j = ocat(ob, j, "\x22,\x22hypothesis\x22:\x22" as *u8); j = ocat(ob, j, axhyp[ai] as *u8)
          j = ocat(ob, j, "\x22,\x22status\x22:\x22HYPOTHESIS-NOT-CLAIM\x22}" as *u8)
          hem = hem + 1
      }
      ai = ai + 1
  }
  j = ocat(ob, j, "]" as *u8)
377
  j = ocat(ob, j, ",\x22axes_total\x22:" as *u8); j = onum(ob, j, NA)
  j = ocat(ob, j, ",\x22reference_titles_total\x22:" as *u8); j = onum(ob, j, NT)
  j = ocat(ob, j, ",\x22cells_total\x22:" as *u8); j = onum(ob, j, ncell)
  j = ocat(ob, j, ",\x22cells_ext\x22:" as *u8); j = onum(ob, j, next)
  j = ocat(ob, j, ",\x22cells_doc\x22:" as *u8); j = onum(ob, j, ndoc)
  j = ocat(ob, j, ",\x22cells_decl\x22:" as *u8); j = onum(ob, j, ndecl)
  j = ocat(ob, j, ",\x22cells_exceeded\x22:" as *u8); j = onum(ob, j, nexceed)
  j = ocat(ob, j, ",\x22beat_index_permil\x22:" as *u8); j = onum(ob, j, beat)
  j = ocat(ob, j, ",\x22beat_verdict\x22:\x22" as *u8); j = ocat(ob, j, bverd); j = ocat(ob, j, "\x22" as *u8)
387
  j = ocat(ob, j, ",\x22honest\x22:\x22beat_index_permil is 0 and the verdict is UNGROUNDED because ZERO head-to-head cells carry EXT evidence. That is not a placeholder, it is the measurement: nobody has ever scored a reference title on our instruments. nx_gamebench publishes the same fact as titles_reference_run=0. Until a cell is EXT, every belief about beating these titles -- including all of the advantage_hypotheses below -- is UNPROVEN, and this organ refuses to convert any of it into a score.\x22" as *u8)
  j = ocat(ob, j, ",\x22why_a_second_queue\x22:\x22nx_gamebench gap_queue ranks by titles_blocked = a PARITY objective (what unblocks the most titles). It cannot express ADVANTAGE: genetics-breeding ranks near the bottom there because only one reference title needs it, while being one of the few axes where a generative emitter beats a hand-authored roster by construction. vsbench ranks the same world by evidence-needed and by advantage class. Two objectives, two queues.\x22" as *u8)
  j = ocat(ob, j, ",\x22reality_anchor\x22:\x22operator 2026-08-09: the current OURS frame ranks below-1980s-commercial class perceptually; a cross-class stat win is Goodhart, so EXCEED additionally requires a SAME-RENDER-CLASS pairing (ccls=0) or a human-confirmed perceptual win. Cross-class stat-ahead cells report MIXED and never count toward the beat index. INSTRUMENT CONVICTION (measured same day): a random-squares mosaic control scores 628 RICH on the frame critic while commercial renders read 789-1000 -- every visual-scene-quality number on this board is a STRUCTURE statistic, not a quality rank.\x22" as *u8)
  j = ocat(ob, j, ",\x22method\x22:\x22EXCEED IS STRICT: mine>theirs=EXCEED, mine==theirs=PARITY, else BEHIND. A tie is never a win. beat_index_permil = 1000*exceeded_ext_cells/ext_cells; undefined (0/UNGROUNDED) when ext_cells=0. Cross-class cells (ccls=1) are excluded from exceeded_ext_cells by construction.\x22" as *u8)
  j = ocat(ob, j, ",\x22gate\x22:\x22nx_vsbench_gate (anti-vacuity: with no EXT evidence the headline MUST be 0/UNGROUNDED; anti-fabrication: every ours_permil names an artifact stat-able on disk)\x22}" as *u8)
  j = ocat(ob, j, "\n" as *u8)
394
  sys_write(1, ob, j)
  return 0
397}
nx_vsbench.nx source

`nx_vsbench.nx` source
