nx_bench_harness_test.nx source
↩ module page · 86 lines · 4824 B
1// nx_bench_harness_test.nx -- KAT for the leaderboard engine.
2// Native sovereign lane; exit 0 = pass, N = assertion N failed.
3//
4// Proves three things on toy-but-exact data:
5// 1. an IDEAL ordering scores the exact reference values (nDCG=MAP=MRR=
6// canonical=1.0; primary precision + S-recall = the closed-form fractions)
7// 2. a JUNK-FIRST ordering (reaction video + listicle on top) is penalized
8// 3. nx_bench_nishi_rank re-orders the SAME documents into a measured
9// S-class EXCEED over the junk ordering (nx_bench_exceeds == 1)
10
11import "fx.nx"
12import "nx_bench_harness.nx"
13
14func main() -> i64 {
15 // ---- gold qrels: 4 first-class docs across facets {0,1,2,4} ----
16 let qn: i64 = 4
17 let qurl: **u8 = sys_mmap(qn * 8) as **u8
18 let qfac: *i64 = sys_mmap(qn * 8) as *i64
19 let qgrd: *i64 = sys_mmap(qn * 8) as *i64
20 let qpri: *i64 = sys_mmap(qn * 8) as *i64
21 qurl[0] = "https://en.wikipedia.org/wiki/Diora_Baird"; qfac[0] = 0; qgrd[0] = 4; qpri[0] = 1
22 qurl[1] = "https://www.imdb.com/title/tt001/"; qfac[1] = 1; qgrd[1] = 4; qpri[1] = 1
23 qurl[2] = "https://www.gettyimages.com/photos/diora"; qfac[2] = 2; qgrd[2] = 3; qpri[2] = 1
24 qurl[3] = "https://variety.com/news/diora-cast"; qfac[3] = 4; qgrd[3] = 3; qpri[3] = 0
25
26 // ---- Test 1: IDEAL ordering (gains 1.0,1.0,1.0,0.9 == sorted desc) ----
27 let n1: i64 = 4
28 let s1u: **u8 = sys_mmap(n1 * 8) as **u8
29 let s1t: **u8 = sys_mmap(n1 * 8) as **u8
30 s1u[0] = "https://en.wikipedia.org/wiki/Diora_Baird"; s1t[0] = "Diora Baird"
31 s1u[1] = "https://www.imdb.com/title/tt001/"; s1t[1] = "Film A (2005)"
32 s1u[2] = "https://www.gettyimages.com/photos/diora"; s1t[2] = "Diora Baird"
33 s1u[3] = "https://variety.com/news/diora-cast"; s1t[3] = "Diora joins film"
34 let sc1: *NxScorecard = sys_mmap(NX_SCORECARD_BYTES) as *NxScorecard
35 nx_bench_score_serp(s1u, s1t, n1, qurl, qfac, qgrd, qpri, qn, 4, sc1)
36 if sc1.ndcg_q16 != FX_ONE { return 1 }
37 if sc1.primary_prec_q16 != fx_from_frac(3, 4) { return 2 } // 3 of 4 first-hand
38 if sc1.s_recall_q16 != fx_from_frac(4, 7) { return 3 } // facets {0,1,2,4}/7
39 if sc1.map_q16 != FX_ONE { return 4 }
40 if sc1.mrr_q16 != FX_ONE { return 5 }
41 if sc1.canonical_rate_q16 != FX_ONE { return 6 }
42
43 // ---- Test 2: JUNK-FIRST (reaction + listicle on top), scored @3 ----
44 let n2: i64 = 6
45 let s2u: **u8 = sys_mmap(n2 * 8) as **u8
46 let s2t: **u8 = sys_mmap(n2 * 8) as **u8
47 s2u[0] = "https://www.youtube.com/watch?v=abc"; s2t[0] = "Diora Baird Hostel REACTION!!"
48 s2u[1] = "https://screenrant.com/diora-baird-movies-ranked/"; s2t[1] = "10 Diora Baird Movies Ranked"
49 s2u[2] = "https://en.wikipedia.org/wiki/Diora_Baird"; s2t[2] = "Diora Baird"
50 s2u[3] = "https://www.imdb.com/title/tt001/"; s2t[3] = "Film A (2005)"
51 s2u[4] = "https://www.gettyimages.com/photos/diora"; s2t[4] = "Diora Baird"
52 s2u[5] = "https://variety.com/news/diora-cast"; s2t[5] = "Diora joins film"
53 let sc2: *NxScorecard = sys_mmap(NX_SCORECARD_BYTES) as *NxScorecard
54 nx_bench_score_serp(s2u, s2t, n2, qurl, qfac, qgrd, qpri, qn, 3, sc2)
55 if sc2.ndcg_q16 >= FX_ONE { return 7 } // junk-first can't be ideal
56 if sc2.primary_prec_q16 != fx_from_frac(1, 3) { return 8 } // only wiki primary in top-3
57
58 // ---- Test 3: Nishi re-rank of the SAME 6 docs -> measured EXCEED ----
59 let pn: i64 = 6
60 let pu: **u8 = sys_mmap(pn * 8) as **u8
61 let pt: **u8 = sys_mmap(pn * 8) as **u8
62 let pf: *i64 = sys_mmap(pn * 8) as *i64
63 pu[0] = s2u[0]; pt[0] = s2t[0]; pf[0] = 0 - 1 // youtube reaction (unjudged junk)
64 pu[1] = s2u[1]; pt[1] = s2t[1]; pf[1] = 0 - 1 // screenrant listicle (unjudged junk)
65 pu[2] = s2u[2]; pt[2] = s2t[2]; pf[2] = 0 // wiki facet 0
66 pu[3] = s2u[3]; pt[3] = s2t[3]; pf[3] = 1 // imdb facet 1
67 pu[4] = s2u[4]; pt[4] = s2t[4]; pf[4] = 2 // getty facet 2
68 pu[5] = s2u[5]; pt[5] = s2t[5]; pf[5] = 4 // variety facet 4
69 let ord: *i64 = sys_mmap(pn * 8) as *i64
70 nx_bench_nishi_rank(pu, pt, pf, pn, ord)
71 let s3u: **u8 = sys_mmap(pn * 8) as **u8
72 let s3t: **u8 = sys_mmap(pn * 8) as **u8
73 var z: i64 = 0
74 while z < pn {
75 s3u[z] = pu[ord[z]]
76 s3t[z] = pt[ord[z]]
77 z = z + 1
78 }
79 let sc3: *NxScorecard = sys_mmap(NX_SCORECARD_BYTES) as *NxScorecard
80 nx_bench_score_serp(s3u, s3t, pn, qurl, qfac, qgrd, qpri, qn, 3, sc3)
81 if sc3.ndcg_q16 != FX_ONE { return 9 } // primaries floated into top-3
82 if sc3.primary_prec_q16 != FX_ONE { return 10 } // top-3 all first-hand
83 if nx_bench_exceeds(sc3, sc2) != 1 { return 11 } // S-class EXCEED, measured
84
85 return 0
86}