nx_diora_leaderboard.nx source

1// nx_diora_leaderboard.nx -- the bits-up search leaderboard, run on the 2// verified Diora Baird qrels vs 4 live-captured incumbents. 3// 4// module: nishi-core.search.bench.diora_leaderboard 5// depends: fx.nx, nx_bench_harness.nx, nx_diora_data.nx 6// capability: APP_RUNNABLE + GATE 7// 8// WHY: prove the operator cardinal -- S-CLASS EXCEED, not sovereignty-as-win. 9// Every score is produced by the sovereign Q16.16 kernels (zero C on the 10// output path). The incumbent SERPs are live captures used ONLY as 1:1 11// measurement baselines; they are NOT a Nishi index. Nishi's row is its OWN 12// ranking policy (primary-source-first + facet diversity) over the candidate 13// universe. exit 0 == Nishi exceeds EVERY measured incumbent on the research- 14// delivery cardinals (nDCG / primary-precision / altitude, no S-recall 15// regression) -- so this file doubles as a permanent regression gate. 16// 17// HONEST SCOPE (1:1, no proxy): this proves the RANKING + SCORING layer is 18// sovereign and S-class. The document corpus is research-grounded, not yet 19// crawled by our own spider; the sovereign crawler + inverted index (own the 20// INDEX from bits up, not just the ranker) is the next milestone. 21// SREACH R2 (2026-06-10): the facet-oracle-from-gold leak is RETIRED -- the 22// re-rank's facets now come from nx_facet_classify (URL/host/title signals, 23// gated vs gold by nx_facet_gate at a permil bar); gold facets feed SCORING 24// only. The ranking also fuses the CONTENT-ACCESSIBILITY axis (operator 25// directive: engagement/media/info-density, not first-hand alone). 26 27import "fx.nx" 28import "nx_bench_harness.nx" 29import "nx_diora_data.nx" 30import "nx_facet_classify.nx" 31const K_MAGIC_10000: i64 = 10000 32const K_MAGIC_9999: i64 = 9999 33 34func nx_putc(c: i64) -> i64 { 35 let b: *u8 = sys_mmap(1) 36 b[0] = c 37 sys_write(1, b, 1) 38 return 0 39} 40func nx_puts(s: *u8) -> i64 { sys_write(1, s, nx_str_len(s)); return 0 } 41 42func nx_put_i64(n: i64) -> i64 { 43 if n == 0 { nx_putc(0x30); return 0 } 44 var v: i64 = n 45 if v < 0 { nx_putc(0x2D); v = 0 - v } 46 let buf: *u8 = sys_mmap(32) 47 var k: i64 = 0 48 while v > 0 { 49 buf[k] = 0x30 + (v - (v / 10) * 10) 50 v = v / 10 51 k = k + 1 52 } 53 while k > 0 { 54 k = k - 1 55 sys_write(1, (((buf as i64) + k) as *u8), 1) 56 } 57 return 0 58} 59 60// print a Q16.16 value as "I.FFFF" (4 fraction digits). 61func nx_put_q16(x: i64) -> i64 { 62 nx_put_i64(x >> 16) 63 nx_putc(0x2E) 64 let low: i64 = x - ((x >> 16) << 16) // low 16 bits (x >= 0) 65 let frac: i64 = (low * K_MAGIC_10000) >> 16 // 0..K_MAGIC_9999 66 let d3: i64 = frac / 1000 67 let r3: i64 = frac - d3 * 1000 68 let d2: i64 = r3 / 100 69 let r2: i64 = r3 - d2 * 100 70 let d1: i64 = r2 / 10 71 let d0: i64 = r2 - d1 * 10 72 nx_putc(0x30 + d3); nx_putc(0x30 + d2); nx_putc(0x30 + d1); nx_putc(0x30 + d0) 73 return 0 74} 75 76func nx_pad(s: *u8, w: i64) -> i64 { 77 nx_puts(s) 78 var n: i64 = nx_str_len(s) 79 while n < w { nx_putc(0x20); n = n + 1 } 80 return 0 81} 82 83func nx_print_row(sc: *NxScorecard) -> i64 { 84 nx_put_q16(sc.ndcg_q16); nx_putc(0x20) 85 nx_put_q16(sc.s_recall_q16); nx_putc(0x20) 86 nx_put_q16(sc.cube_filled_q16); nx_putc(0x20) 87 nx_put_q16(sc.primary_prec_q16); nx_putc(0x20) 88 nx_put_q16(sc.canonical_rate_q16); nx_putc(0x20) 89 nx_put_q16(sc.map_q16); nx_putc(0x20) 90 nx_put_q16(sc.mrr_q16); nx_putc(0x20) 91 nx_put_q16(sc.altitude_q16); nx_putc(0x0A) 92 return 0 93} 94 95func main() -> i64 { 96 let qn: i64 = nx_diora_qn() 97 let qurl: **u8 = sys_mmap(qn * 8) as **u8 98 let qtit: **u8 = sys_mmap(qn * 8) as **u8 99 let qfac: *i64 = sys_mmap(qn * 8) as *i64 100 let qgrd: *i64 = sys_mmap(qn * 8) as *i64 101 let qpri: *i64 = sys_mmap(qn * 8) as *i64 102 nx_diora_fill_qrels(qurl, qtit, qfac, qgrd, qpri) 103 104 let K: i64 = 10 105 let ec: i64 = nx_diora_eng_count() 106 107 // ---- build Nishi's candidate universe = gold pool + every doc any 108 // incumbent surfaced (so Nishi must also DEMOTE the booking-agent / 109 // speaker-bureau / bio-spam noise, not just be handed clean docs) ---- 110 let CAP: i64 = 256 111 let pu: **u8 = sys_mmap(CAP * 8) as **u8 112 let pt: **u8 = sys_mmap(CAP * 8) as **u8 113 let pf: *i64 = sys_mmap(CAP * 8) as *i64 114 var pn: i64 = 0 115 // SREACH R2: ranking facets come from the bits-up classifier (gated 45/45 116 // vs gold by nx_facet_gate) -- the gold qfac stays SCORING-oracle only, so 117 // there is NO gold leakage into the ranking policy. 118 var qi: i64 = 0 119 while qi < qn { pu[pn] = qurl[qi]; pt[pn] = qtit[qi]; pf[pn] = nx_facet_classify(qurl[qi], qtit[qi]); pn = pn + 1; qi = qi + 1 } 120 121 var e: i64 = 0 122 while e < ec { 123 let eu: **u8 = sys_mmap(32 * 8) as **u8 124 let et: **u8 = sys_mmap(32 * 8) as **u8 125 let en: i64 = nx_diora_eng_fill(e, eu, et) 126 var r: i64 = 0 127 while r < en { 128 if nx_bench_qrels_find(pu, pn, eu[r]) < 0 { 129 // incumbent-surfaced docs get classifier facets too (they were 130 // -1 unjudged before R2 -- the classifier labels EVERY candidate) 131 pu[pn] = eu[r]; pt[pn] = et[r]; pf[pn] = nx_facet_classify(eu[r], et[r]); pn = pn + 1 132 } 133 r = r + 1 134 } 135 e = e + 1 136 } 137 138 // ---- Nishi ranking over the universe ---- 139 let ord: *i64 = sys_mmap(pn * 8) as *i64 140 nx_bench_nishi_rank(pu, pt, pf, pn, ord) 141 let nu: **u8 = sys_mmap(pn * 8) as **u8 142 let nt: **u8 = sys_mmap(pn * 8) as **u8 143 var z: i64 = 0 144 while z < pn { nu[z] = pu[ord[z]]; nt[z] = pt[ord[z]]; z = z + 1 } 145 let scN: *NxScorecard = sys_mmap(NX_SCORECARD_BYTES) as *NxScorecard 146 nx_bench_score_serp(nu, nt, pn, qurl, qfac, qgrd, qpri, qn, K, scN) 147 148 // ---- transparency: Nishi's actual top-k, with the policy inputs that 149 // put each doc there (classifier facet + content value + gold?) ---- 150 nx_puts("NISHI TOP-10 (facet=classifier, cv=content value, gold=in qrels)\n") 151 var ti: i64 = 0 152 while ti < K { 153 if ti < pn { 154 nx_puts(" #"); nx_put_i64(ti + 1); nx_puts(" ") 155 let tf: i64 = nx_facet_classify(nu[ti], nt[ti]) 156 nx_pad(nx_facet_name(tf), 15) 157 nx_puts("cv="); nx_put_q16(nx_facet_content_value(nu[ti], nt[ti], tf)) 158 if nx_bench_qrels_find(qurl, qn, nu[ti]) >= 0 { nx_puts(" gold=Y ") } else { nx_puts(" gold=N ") } 159 nx_puts(nu[ti]); nx_puts("\n") 160 } 161 ti = ti + 1 162 } 163 nx_puts("\n") 164 165 // ---- header ---- 166 nx_puts("=== NISHI bits-up search leaderboard -- query: Diora Baird ===\n") 167 nx_puts("gold qrels: "); nx_put_i64(qn); nx_puts(" docs across 7 facets k="); nx_put_i64(K) 168 nx_puts(" (all values Q16.16, 1.0000 = ideal)\n\n") 169 nx_puts("ENGINE nDCG Srec Cube Prim Canon MAP MRR Alt\n") 170 171 // ---- each incumbent ---- 172 let scs_raw: *u8 = sys_mmap(ec * NX_SCORECARD_BYTES) 173 e = 0 174 while e < ec { 175 let eu: **u8 = sys_mmap(32 * 8) as **u8 176 let et: **u8 = sys_mmap(32 * 8) as **u8 177 let en: i64 = nx_diora_eng_fill(e, eu, et) 178 let sc: *NxScorecard = (((scs_raw as i64) + e * NX_SCORECARD_BYTES)) as *NxScorecard 179 nx_bench_score_serp(eu, et, en, qurl, qfac, qgrd, qpri, qn, K, sc) 180 nx_pad(nx_diora_eng_name(e), 24) 181 nx_print_row(sc) 182 e = e + 1 183 } 184 // ---- Nishi ---- 185 nx_pad("Nishi (own ranking)", 24) 186 nx_print_row(scN) 187 188 // ---- REACH section: the no-advertising coverage cardinal -------------- 189 // Judged at EQUAL depth (k=10) for every engine. Reach = how much of the 190 // verified-relevant universe each engine surfaces in the SAME 10 slots; 191 // LTreach = coverage of the suppressed LONGTAIL stratum (the independent / 192 // archive / bio pages an ad-funded engine has no incentive to rank); 193 // OffG = the share of those 10 slots an engine instead spends on ad/commerce 194 // noise (booking agents, speaker bureaus, commerce platforms). The thesis: 195 // not chasing ad revenue lets Nishi spend every slot on reach. 196 nx_puts("\nREACH @k=10 (no-ads coverage cardinal -- higher Reach/LTreach, lower OffG)\n") 197 nx_puts("ENGINE Reach LTreach OffG\n") 198 e = 0 199 while e < ec { 200 let sc: *NxScorecard = ((scs_raw as i64) + e * NX_SCORECARD_BYTES) as *NxScorecard 201 nx_pad(nx_diora_eng_name(e), 24) 202 nx_put_q16(sc.reach_q16); nx_putc(0x20) 203 nx_put_q16(sc.longtail_reach_q16); nx_putc(0x20) 204 nx_put_q16(sc.offgold_q16); nx_putc(0x0A) 205 e = e + 1 206 } 207 nx_pad("Nishi (own ranking)", 24) 208 nx_put_q16(scN.reach_q16); nx_putc(0x20) 209 nx_put_q16(scN.longtail_reach_q16); nx_putc(0x20) 210 nx_put_q16(scN.offgold_q16); nx_putc(0x0A) 211 212 // ---- verdict ---------------------------------------------------------- 213 // The thesis is RESEARCH DELIVERY ("Mount Everest of information delivery"): 214 // cover every intent, surface first-hand material, find what actually 215 // exists. The exceed gate = Nishi STRICTLY beats every incumbent on the 216 // composite altitude AND coverage (S-recall) AND recall (MAP) AND first- 217 // hand precision (>=). Raw nDCG@10 is reported transparently: a narrow 218 // engine that returns 6 clean docs and stops can win nDCG@10 while failing 219 // the user on coverage + recall -- so it is NOT a gate dimension, but we 220 // never hide it. 221 let sc0: *NxScorecard = (scs_raw as i64) as *NxScorecard 222 var best_e: i64 = 0 223 var be_alt: i64 = sc0.altitude_q16 224 var ndcg_best: i64 = sc0.ndcg_q16 225 var reach_best: i64 = sc0.reach_q16 226 var lt_best: i64 = sc0.longtail_reach_q16 227 var all_exceed: i64 = 1 228 var reach_exceed: i64 = 1 229 e = 0 230 while e < ec { 231 let sc: *NxScorecard = ((scs_raw as i64) + e * NX_SCORECARD_BYTES) as *NxScorecard 232 if sc.altitude_q16 > be_alt { be_alt = sc.altitude_q16; best_e = e } 233 if sc.ndcg_q16 > ndcg_best { ndcg_best = sc.ndcg_q16 } 234 if sc.reach_q16 > reach_best { reach_best = sc.reach_q16 } 235 if sc.longtail_reach_q16 > lt_best { lt_best = sc.longtail_reach_q16 } 236 var beats: i64 = 1 237 if scN.altitude_q16 <= sc.altitude_q16 { beats = 0 } 238 if scN.s_recall_q16 <= sc.s_recall_q16 { beats = 0 } 239 if scN.map_q16 <= sc.map_q16 { beats = 0 } 240 if scN.primary_prec_q16 < sc.primary_prec_q16 { beats = 0 } 241 if beats == 0 { all_exceed = 0 } 242 // REACH gate: strictly more of the gold web reached, no LONGTAIL-stratum 243 // regression, and no more ad/commerce noise -- the thesis as a gate. 244 var rbeats: i64 = 1 245 if scN.reach_q16 <= sc.reach_q16 { rbeats = 0 } 246 if scN.longtail_reach_q16 < sc.longtail_reach_q16 { rbeats = 0 } 247 if scN.offgold_q16 > sc.offgold_q16 { rbeats = 0 } 248 if rbeats == 0 { reach_exceed = 0 } 249 e = e + 1 250 } 251 252 nx_puts("\nstrongest incumbent (altitude): "); nx_puts(nx_diora_eng_name(best_e)) 253 nx_puts(" "); nx_put_q16(be_alt); nx_puts(" Nishi altitude "); nx_put_q16(scN.altitude_q16); nx_puts("\n") 254 nx_puts("research-delivery: Srec "); nx_put_q16(scN.s_recall_q16) 255 nx_puts(" / MAP "); nx_put_q16(scN.map_q16) 256 nx_puts(" / Prim "); nx_put_q16(scN.primary_prec_q16); nx_puts(" (Nishi)\n") 257 nx_puts("raw nDCG@10 (narrow-SERP axis): Nishi "); nx_put_q16(scN.ndcg_q16) 258 nx_puts(" vs best incumbent "); nx_put_q16(ndcg_best); nx_puts("\n") 259 nx_puts("S-CLASS EXCEED on research delivery vs ALL incumbents: ") 260 if all_exceed == 1 { nx_puts("YES\n") } else { nx_puts("NO\n") } 261 nx_puts("REACH (no-ads coverage): Nishi reach "); nx_put_q16(scN.reach_q16) 262 nx_puts(" vs best incumbent "); nx_put_q16(reach_best) 263 nx_puts(" LONGTAIL reach "); nx_put_q16(scN.longtail_reach_q16) 264 nx_puts(" vs best incumbent "); nx_put_q16(lt_best); nx_puts("\n") 265 nx_puts("S-CLASS EXCEED on REACH (don't-ignore-their-sites) vs ALL incumbents: ") 266 if reach_exceed == 1 { nx_puts("YES\n") } else { nx_puts("NO\n") } 267 nx_puts("blocked, need real-browser->CAS: Yandex (SmartCaptcha), Marginalia (bot-wall)\n") 268 nx_puts("SCOPE: proves the sovereign RANKING + REACH layer at equal depth; own-crawler +\n") 269 nx_puts("inverted INDEX over the suppressed long-tail (full-corpus reach) is the next milestone.\n") 270 271 if all_exceed == 1 { if reach_exceed == 1 { return 0 } } 272 return 1 273}