nx_source_tier.nx source

1// nx_source_tier.nx -- the research-established ranking rule, bits-up. 2// 3// module: nishi-core.search.bench.source_tier 4// depends: fx.nx (Q16.16 tier scores), nx_str.nx (substring matching) 5// capability: CORE_COMPUTE 6// wired_status: FULLY_WIRED 7// 8// Operationalizes the operator's principle: FIRST-HAND source material (the 9// actual film / official photo / canonical record / direct interview) ranks 10// ABOVE derivative commentary (reaction videos, "top 10" listicles, SEO 11// aggregators, tier-lists). This is the anti-Google ranking thesis from 12// docs/00-VISION + the five-tier taxonomy from docs/04-DEDUP. 13// 14// Two-stage, data-driven (Rule #11 -- the tables ARE the policy): 15// 1. domain -> base tier (primary/editorial/platform/archive/community) 16// 2. title/url "derivative" signals DEMOTE a result to commentary even on 17// a high-tier domain (a YouTube reaction video is platform AND 18// derivative; the same channel's official trailer is not). 19// 20// A result is_primary (first-hand) iff base tier == PRIMARY and no derivative 21// signal fired. That boolean feeds nx_bench_primary_precision_at_k; the gain 22// (tier score, zeroed when derivative) feeds nx_bench_dcg/ndcg. 23// 24// Tier scores from docs/04-DEDUP: primary 1.0, editorial 0.9, platform 0.7, 25// archive 0.6, community 0.5, unknown 0.3 (Q16.16). 26 27import "fx.nx" 28import "nx_str.nx" 29 30const NX_TIER_PRIMARY: i64 = 1 31const NX_TIER_EDITORIAL: i64 = 2 32const NX_TIER_PLATFORM: i64 = 3 33const NX_TIER_ARCHIVE: i64 = 4 34const NX_TIER_COMMUNITY: i64 = 5 35const NX_TIER_UNKNOWN: i64 = 6 36 37struct NxTierVerdict { 38 tier: i64, 39 score: i64, // Q16.16 tier score 40 is_primary: i64, // 1 iff first-hand source 41 is_derivative: i64, // 1 iff commentary/aggregation about the work 42} 43 44const NX_TIER_VERDICT_BYTES: i64 = 32 45 46// Q16.16 tier score for a tier id. 47func nx_tier_score(tier: i64) -> i64 { 48 if tier == NX_TIER_PRIMARY { return FX_ONE } 49 if tier == NX_TIER_EDITORIAL { return fx_from_frac(9, 10) } 50 if tier == NX_TIER_PLATFORM { return fx_from_frac(7, 10) } 51 if tier == NX_TIER_ARCHIVE { return fx_from_frac(6, 10) } 52 if tier == NX_TIER_COMMUNITY { return fx_from_frac(1, 2) } 53 return fx_from_frac(3, 10) 54} 55 56// ASCII lowercase of a single byte value. 57func nx_tier_lc(c: i64) -> i64 { 58 if c >= 0x41 { 59 if c <= 0x5A { return c + 0x20 } 60 } 61 return c 62} 63 64// Case-insensitive "does hay contain needle". needle must be lowercase. 65// No scratch buffer / no byte-store cast: folds case during comparison. 66func nx_tier_ci_contains(hay: *u8, needle: *u8) -> i64 { 67 let hn: i64 = nx_str_len(hay) 68 let nn: i64 = nx_str_len(needle) 69 if nn == 0 { return 1 } 70 if nn > hn { return 0 } 71 var i: i64 = 0 72 while i <= hn - nn { 73 var j: i64 = 0 74 var ok: i64 = 1 75 while j < nn { 76 if nx_tier_lc(hay[i + j] as i64) != nx_tier_lc(needle[j] as i64) { 77 ok = 0 78 j = nn 79 } else { 80 j = j + 1 81 } 82 } 83 if ok == 1 { return 1 } 84 i = i + 1 85 } 86 return 0 87} 88 89// host substring test (domains are lowercase in URLs -> case-sensitive ok). 90func nx_tier_host_has(url: *u8, domain: *u8) -> i64 { 91 if nx_str_str(url, domain) >= 0 { return 1 } 92 return 0 93} 94 95// domain -> base tier. One line per domain (Rule #11 data-driven). Archive 96// checked before generic so web.archive.org is archive, not unknown. 97func nx_tier_of_url(url: *u8) -> i64 { 98 // primary: the canonical record / official / the work or photo itself 99 if nx_tier_host_has(url, "imdb.com") == 1 { return NX_TIER_PRIMARY } 100 if nx_tier_host_has(url, "wikipedia.org") == 1 { return NX_TIER_PRIMARY } 101 if nx_tier_host_has(url, "wikidata.org") == 1 { return NX_TIER_PRIMARY } 102 if nx_tier_host_has(url, "wikimedia.org") == 1 { return NX_TIER_PRIMARY } 103 if nx_tier_host_has(url, "gettyimages.com") == 1 { return NX_TIER_PRIMARY } 104 if nx_tier_host_has(url, "apimages.com") == 1 { return NX_TIER_PRIMARY } 105 if nx_tier_host_has(url, "arxiv.org") == 1 { return NX_TIER_PRIMARY } 106 if nx_tier_host_has(url, "openreview.net") == 1 { return NX_TIER_PRIMARY } 107 // NOTE (verified 2026-05-29, design workflow): IMDb + Wikipedia are the 108 // only reliable PRIMARY anchors for a celebrity query. Rotten Tomatoes, 109 // TMDB and Box Office Mojo are derivative aggregator CATALOGS that 110 // dominate mainstream SERPs -- they are secondary, not first-hand. Tiered 111 // editorial below so they never earn primary gain. 112 // archive: Wayback / dead-host recovery 113 if nx_tier_host_has(url, "web.archive.org") == 1 { return NX_TIER_ARCHIVE } 114 if nx_tier_host_has(url, "archive.org") == 1 { return NX_TIER_ARCHIVE } 115 if nx_tier_host_has(url, "archive.ph") == 1 { return NX_TIER_ARCHIVE } 116 // editorial: professional press + derivative aggregator catalogs 117 if nx_tier_host_has(url, "themoviedb.org") == 1 { return NX_TIER_EDITORIAL } 118 if nx_tier_host_has(url, "rottentomatoes.com") == 1 { return NX_TIER_EDITORIAL } 119 if nx_tier_host_has(url, "boxofficemojo.com") == 1 { return NX_TIER_EDITORIAL } 120 if nx_tier_host_has(url, "filmaffinity.com") == 1 { return NX_TIER_EDITORIAL } 121 if nx_tier_host_has(url, "tvguide.com") == 1 { return NX_TIER_EDITORIAL } 122 if nx_tier_host_has(url, "variety.com") == 1 { return NX_TIER_EDITORIAL } 123 if nx_tier_host_has(url, "hollywoodreporter.com") == 1 { return NX_TIER_EDITORIAL } 124 if nx_tier_host_has(url, "ew.com") == 1 { return NX_TIER_EDITORIAL } 125 if nx_tier_host_has(url, "people.com") == 1 { return NX_TIER_EDITORIAL } 126 if nx_tier_host_has(url, "nytimes.com") == 1 { return NX_TIER_EDITORIAL } 127 if nx_tier_host_has(url, "vanityfair.com") == 1 { return NX_TIER_EDITORIAL } 128 if nx_tier_host_has(url, "maxim.com") == 1 { return NX_TIER_EDITORIAL } 129 if nx_tier_host_has(url, "deadline.com") == 1 { return NX_TIER_EDITORIAL } 130 if nx_tier_host_has(url, "screenrant.com") == 1 { return NX_TIER_EDITORIAL } 131 if nx_tier_host_has(url, "collider.com") == 1 { return NX_TIER_EDITORIAL } 132 if nx_tier_host_has(url, "cbr.com") == 1 { return NX_TIER_EDITORIAL } 133 // verified-real editorial/press + first-hand-interview outlets (2026-05-29): 134 // these are genuine publications, not SEO bio-spam, so they belong in the 135 // editorial table (Rule #11) -- a person query's interviews/press facets 136 // live here, not in "unknown". 137 if nx_tier_host_has(url, "advocate.com") == 1 { return NX_TIER_EDITORIAL } 138 if nx_tier_host_has(url, "losangelesblade.com") == 1 { return NX_TIER_EDITORIAL } 139 if nx_tier_host_has(url, "sbs.com.au") == 1 { return NX_TIER_EDITORIAL } 140 if nx_tier_host_has(url, "bloody-disgusting.com") == 1 { return NX_TIER_EDITORIAL } 141 if nx_tier_host_has(url, "naludamagazine.com") == 1 { return NX_TIER_EDITORIAL } 142 if nx_tier_host_has(url, "pop-culturalist.com") == 1 { return NX_TIER_EDITORIAL } 143 if nx_tier_host_has(url, "afterbuzztv.com") == 1 { return NX_TIER_EDITORIAL } 144 if nx_tier_host_has(url, "dontgooutthere.com") == 1 { return NX_TIER_EDITORIAL } 145 // platform: user-content platforms 146 if nx_tier_host_has(url, "youtube.com") == 1 { return NX_TIER_PLATFORM } 147 if nx_tier_host_has(url, "youtu.be") == 1 { return NX_TIER_PLATFORM } 148 if nx_tier_host_has(url, "reddit.com") == 1 { return NX_TIER_PLATFORM } 149 if nx_tier_host_has(url, "twitter.com") == 1 { return NX_TIER_PLATFORM } 150 if nx_tier_host_has(url, "instagram.com") == 1 { return NX_TIER_PLATFORM } 151 if nx_tier_host_has(url, "tumblr.com") == 1 { return NX_TIER_PLATFORM } 152 if nx_tier_host_has(url, "tiktok.com") == 1 { return NX_TIER_PLATFORM } 153 if nx_tier_host_has(url, "facebook.com") == 1 { return NX_TIER_PLATFORM } 154 if nx_tier_host_has(url, "pinterest.com") == 1 { return NX_TIER_PLATFORM } 155 if nx_tier_host_has(url, "imgur.com") == 1 { return NX_TIER_PLATFORM } 156 if nx_tier_host_has(url, "github.com") == 1 { return NX_TIER_PLATFORM } 157 if nx_tier_host_has(url, "huggingface.co") == 1 { return NX_TIER_PLATFORM } 158 // community: fan wikis / forums 159 if nx_tier_host_has(url, "fandom.com") == 1 { return NX_TIER_COMMUNITY } 160 if nx_tier_host_has(url, "wikia.com") == 1 { return NX_TIER_COMMUNITY } 161 if nx_tier_host_has(url, "proboards.com") == 1 { return NX_TIER_COMMUNITY } 162 return NX_TIER_UNKNOWN 163} 164 165// derivative-commentary signal in title OR url path (lowercase needles). 166func nx_tier_is_derivative(url: *u8, title: *u8) -> i64 { 167 if nx_tier_ci_contains(title, "reaction") == 1 { return 1 } 168 if nx_tier_ci_contains(title, "reacts") == 1 { return 1 } 169 if nx_tier_ci_contains(title, "review") == 1 { return 1 } 170 if nx_tier_ci_contains(title, "breakdown") == 1 { return 1 } 171 if nx_tier_ci_contains(title, "explained") == 1 { return 1 } 172 if nx_tier_ci_contains(title, "top 10") == 1 { return 1 } 173 if nx_tier_ci_contains(title, "top 5") == 1 { return 1 } 174 if nx_tier_ci_contains(title, "ranked") == 1 { return 1 } 175 if nx_tier_ci_contains(title, "ranking") == 1 { return 1 } 176 if nx_tier_ci_contains(title, "tier list") == 1 { return 1 } 177 if nx_tier_ci_contains(title, "compilation") == 1 { return 1 } 178 if nx_tier_ci_contains(title, "hottest") == 1 { return 1 } 179 if nx_tier_ci_contains(title, "sexiest") == 1 { return 1 } 180 if nx_tier_ci_contains(title, "net worth") == 1 { return 1 } 181 if nx_tier_ci_contains(title, "things you") == 1 { return 1 } 182 if nx_tier_ci_contains(title, "facts about") == 1 { return 1 } 183 if nx_tier_ci_contains(url, "ranked") == 1 { return 1 } 184 if nx_tier_ci_contains(url, "-best-") == 1 { return 1 } 185 // NOTE (operator cardinal 2026-05-29): this flag is METADATA ONLY -- a 186 // "looks like commentary/aggregation" hint. It MUST NOT be used to zero a 187 // result's value or prejudge it: a commentary/gallery/compilation page can 188 // carry UNIQUE content (Comic-Con pics, rare footage, original reporting) 189 // and our mission is to DELIVER that information, not suppress it. Earned 190 // demotion of genuinely non-additive duplicates is the novelty/dedup 191 // kernel's job, measured -- never assumed from a title keyword. 192 return 0 193} 194 195// Full classification into a caller-supplied verdict slot. 196func nx_tier_classify(url: *u8, title: *u8, out: *NxTierVerdict) -> i64 { 197 let t: i64 = nx_tier_of_url(url) 198 let deriv: i64 = nx_tier_is_derivative(url, title) 199 out.tier = t 200 out.score = nx_tier_score(t) 201 out.is_derivative = deriv 202 var prim: i64 = 0 203 if t == NX_TIER_PRIMARY { 204 if deriv == 0 { prim = 1 } 205 } 206 out.is_primary = prim 207 return 0 208} 209 210// Ranking-prior gain for a verdict = the tier score. We do NOT zero 211// "derivative" results (operator cardinal 2026-05-29: deliver information, do 212// not prejudge -- a commentary/aggregation/gallery page can carry unique 213// content). is_derivative stays on the verdict as METADATA only; genuine 214// non-additive duplicates are demoted later by the novelty/dedup mechanism, 215// measured. The primary-source PRIOR still ranks first-hand sources highest. 216func nx_tier_gain(v: *NxTierVerdict) -> i64 { 217 return v.score 218}