`nx_dr_fuse.nx` source

↩ module page · 122 lines · 4742 B
1// nx_dr_fuse.nx -- LEARNED signal combiner for the deep-research judge (DR-12).
2// Two hand-reasoned refinements to the count model were measured and REJECTED (a domain-
3// augmented vocabulary made discrimination worse; IDF weighting was a dead heat). What had
4// never been tried is LEARNING how to combine the signals we already compute, from the real
5// labels ServiceNow publishes.
6//
7// Model: score = w0*lexical + w1*semantic + w2*semantic_idf  (integer, deterministic).
8// Fitting: EXHAUSTIVE GRID SEARCH over integer weights on a TRAINING task range, scored by
9// pairwise ranking accuracy (every insight must outrank every distractor within its own task,
10// which is exactly the matched-precision metric the judge is graded on). Grid search rather
11// than gradient descent: no learning rate, no convergence failure, and bit-reproducible.
12//
13// ★HONEST EVALUATION BY CONSTRUCTION: weights are fitted ONLY on the training task range and
14// reported ONLY on a DISJOINT held-out range. Fitting and scoring on the same tasks would be
15// the rigged-gate sin; the split is a parameter so the gate can prove the separation.
16// Every func <=6 params (NAS nx_cc >6-arg skew, seq239). No hardware writes (Rule 26).
17//
18// module: nishi-core.research.dr_fuse
19// depends: nx_syscalls.nx
20// genealogy_id: linear_rank_fusion + perceptron_ranking
21import "nx_syscalls.nx"
22const FU_MAGIC_1000000000: i64 = 1000000000
23
24// record layout: 5 i64 per candidate -- [task, label(1=insight), lexical, semantic, semantic_idf]
25const FU_STRIDE: i64 = 5
26
27func fu_task(r: *i64, i: i64) -> i64 { return r[i * FU_STRIDE] }
28func fu_label(r: *i64, i: i64) -> i64 { return r[i * FU_STRIDE + 1] }
29
30// combined score under weights w[3]
31func fu_score(r: *i64, i: i64, w: *i64) -> i64 {
32    let b: i64 = i * FU_STRIDE
33    return w[0] * r[b + 2] + w[1] * r[b + 3] + w[2] * r[b + 4]
34}
35
36// pairwise ranking accuracy (permille) over tasks in [lo,hi): of all (insight,distractor)
37// pairs inside a task, how many does `w` order correctly?
38func fu_pairacc(r: *i64, n: i64, w: *i64, lo: i64, hi: i64) -> i64 {
39    var good: i64 = 0
40    var tot: i64 = 0
41    var i: i64 = 0
42    while i < n {
43        let t: i64 = fu_task(r, i)
44        if t >= lo { if t < hi { if fu_label(r, i) == 1 {
45            let si: i64 = fu_score(r, i, w)
46            var j: i64 = 0
47            while j < n {
48                if fu_task(r, j) == t { if fu_label(r, j) == 0 {
49                    tot = tot + 1
50                    if si > fu_score(r, j, w) { good = good + 1 }
51                } }
52                j = j + 1
53            }
54        } } }
55        i = i + 1
56    }
57    if tot < 1 { return 0 }
58    return (good * 1000) / tot
59}
60
61// mean insight-recall at 100% distractor-avoidance over tasks in [lo,hi): per task, threshold
62// just above the highest-scoring distractor, then count insights strictly above it.
63func fu_recall(r: *i64, n: i64, w: *i64, lo: i64, hi: i64) -> i64 {
64    var sum: i64 = 0
65    var tasks: i64 = 0
66    var t: i64 = lo
67    while t < hi {
68        var maxd: i64 = 0 - FU_MAGIC_1000000000
69        var ni: i64 = 0
70        var seen: i64 = 0
71        var i: i64 = 0
72        while i < n {
73            if fu_task(r, i) == t {
74                seen = 1
75                if fu_label(r, i) == 0 { let s: i64 = fu_score(r, i, w); if s > maxd { maxd = s } }
76                else { ni = ni + 1 }
77            }
78            i = i + 1
79        }
80        if seen == 1 { if ni > 0 {
81            var hit: i64 = 0
82            i = 0
83            while i < n {
84                if fu_task(r, i) == t { if fu_label(r, i) == 1 {
85                    if fu_score(r, i, w) > maxd { hit = hit + 1 }
86                } }
87                i = i + 1
88            }
89            sum = sum + (hit * 1000) / ni
90            tasks = tasks + 1
91        } }
92        t = t + 1
93    }
94    if tasks < 1 { return 0 }
95    return sum / tasks
96}
97
98// Fit w[3] by exhaustive integer grid search on tasks [lo,hi), maximising pairwise accuracy.
99// Deterministic: ties keep the FIRST (lowest) weight vector, so the result is reproducible.
100// Returns the best training pairwise accuracy achieved.
101func fu_fit(r: *i64, n: i64, w: *i64, lo: i64, hi: i64) -> i64 {
102    let cand: *i64 = sys_mmap(3 * 8) as *i64
103    var best: i64 = 0 - 1
104    var a: i64 = 0
105    while a <= 10 {
106        var b: i64 = 0
107        while b <= 10 {
108            var c: i64 = 0
109            while c <= 10 {
110                if a + b + c > 0 {
111                    cand[0] = a; cand[1] = b; cand[2] = c
112                    let acc: i64 = fu_pairacc(r, n, cand, lo, hi)
113                    if acc > best { best = acc; w[0] = a; w[1] = b; w[2] = c }
114                }
115                c = c + 1
116            }
117            b = b + 1
118        }
119        a = a + 1
120    }
121    return best
122}
nx_dr_fuse.nx source

`nx_dr_fuse.nx` source
