nx_code_embed_gate.nx source
↩ module page · 265 lines · 10373 B
1// nx_code_embed_gate.nx -- SOVEREIGN CODE-EMBEDDING gate (2026-07-14). Grounds the jina-code-embeddings
2// recipe (arXiv:2508.21290, backbone = Qwen2.5-Coder-0.5B = OUR coder) on the no-float stack: embedding =
3// last-token pooled final-norm hidden state = nsv_embed(). Task instruction prefixes = jina's EXACT strings.
4// GATES (liar-killer style):
5// A ORDER jina model-card 2x2 demo: cos(q1,p1)>cos(q1,p2) AND cos(q2,p2)>cos(q2,p1)
6// (bf16 jina-tuned oracle values 817/124/120/553 pm -- we gate ORDER only: base weights differ)
7// B DETERM re-embed q1 AFTER all other texts -> vector must be BITWISE IDENTICAL (catches KV/state leak)
8// C SELF cos(q1,q1') >= 999 pm (isqrt floor can give 1000+/-1)
9// D RETRIEVE 3x3 Nishi-corpus NL->code argmax must hit the diagonal 3/3
10// E API synthetic HTTP POST /embed through nsv_handle -> 200 + ok:1 + dim:896 (the daemon serves
11// these exact bytes -> the capability is API-reachable, rule-27 style, not gate-only)
12// HONEST SCOPE: this proves PIPELINE CORRECTNESS on BASE coder weights. SOTA retrieval quality is NOT
13// claimed -- that grading pends the jina-weight oracle (cc-by-nc-4.0 -> benchmark-only, like gcc).
14// license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0
15import "nx_nofloat_serve_core.nx"
16
17const CE_NE: i64 = 896
18const CE_SLOTS: i64 = 11
19
20func ce_ws(s: *u8) -> i64 {
21 var n: i64 = 0
22 while s[n] != (0 as u8) { n = n + 1 }
23 sys_write(1, s, n)
24 return 0
25}
26
27func ce_wdec(v: i64) -> i64 {
28 let b: *u8 = sys_mmap(28)
29 let t: *u8 = sys_mmap(28)
30 var m: i64 = v
31 if m < 0 { ce_ws("-" as *u8); m = 0 - m }
32 var k: i64 = 0
33 if m == 0 { t[0] = 48 as u8; k = 1 }
34 while m > 0 { t[k] = (48 + (m % 10)) as u8; m = m / 10; k = k + 1 }
35 var i: i64 = 0
36 while i < k { b[i] = t[k - 1 - i]; i = i + 1 }
37 sys_write(1, b, k)
38 return 0
39}
40
41func ce_kv(label: *u8, v: i64) -> i64 {
42 ce_ws(label)
43 ce_ws("=" as *u8)
44 ce_wdec(v)
45 ce_ws(" " as *u8)
46 return 0
47}
48
49func ce_isqrt(v: i64) -> i64 {
50 if v < 2 { return v }
51 var x: i64 = v
52 var y: i64 = (x + 1) / 2
53 while y < x { x = y; y = (x + v / x) / 2 }
54 return x
55}
56
57// cosine similarity in per-mille, overflow-safe: pre-scale both vectors by the SAME power of two so
58// max|elem| <= 2^20 -> squares*ne <= 2^50, dot*1000 <= 2^60, isqrt product <= 2^50 (all inside i64).
59func ce_cos_pm(a: *i64, b: *i64) -> i64 {
60 var mx: i64 = 1
61 var i: i64 = 0
62 while i < CE_NE {
63 var va: i64 = a[i]
64 if va < 0 { va = 0 - va }
65 var vb: i64 = b[i]
66 if vb < 0 { vb = 0 - vb }
67 if va > mx { mx = va }
68 if vb > mx { mx = vb }
69 i = i + 1
70 }
71 var dv: i64 = 1
72 while mx > 1048575 { mx = mx / 2; dv = dv * 2 }
73 var dab: i64 = 0
74 var daa: i64 = 0
75 var dbb: i64 = 0
76 i = 0
77 while i < CE_NE {
78 let ea: i64 = a[i] / dv
79 let eb: i64 = b[i] / dv
80 dab = dab + ea * eb
81 daa = daa + ea * ea
82 dbb = dbb + eb * eb
83 i = i + 1
84 }
85 let sa: i64 = ce_isqrt(daa)
86 let sb: i64 = ce_isqrt(dbb)
87 if sa == 0 { return 0 }
88 if sb == 0 { return 0 }
89 return (dab * 1000) / (sa * sb)
90}
91
92func ce_bytes_eq(a: *i64, b: *i64) -> i64 {
93 var i: i64 = 0
94 while i < CE_NE { if a[i] != b[i] { return 0 } i = i + 1 }
95 return 1
96}
97
98// substring search: index of needle in hay[0,hn) or -1
99func ce_find(hay: *u8, hn: i64, needle: *u8) -> i64 {
100 var nl: i64 = 0
101 while needle[nl] != (0 as u8) { nl = nl + 1 }
102 if nl == 0 { return 0 }
103 var i: i64 = 0
104 while i + nl <= hn {
105 var k: i64 = 0
106 var ok: i64 = 1
107 while k < nl { if hay[i+k] != needle[k] { ok = 0; k = nl } else { k = k + 1 } }
108 if ok == 1 { return i }
109 i = i + 1
110 }
111 return 0 - 1
112}
113
114func ce_maxabs(a: *i64) -> i64 {
115 var mx: i64 = 0
116 var i: i64 = 0
117 while i < CE_NE {
118 var v: i64 = a[i]
119 if v < 0 { v = 0 - v }
120 if v > mx { mx = v }
121 i = i + 1
122 }
123 return mx
124}
125
126// embed prefix+content into vec slot; prints slot/ntok/ms; returns ntok (negative = error)
127func ce_embed(vecs: *i64, slot: i64, pre: *u8, content: *u8) -> i64 {
128 let txt: *u8 = sys_mmap(2048)
129 var l: i64 = nsv_cat(txt, 0, pre)
130 l = nsv_cat(txt, l, content)
131 let vp_i: i64 = (vecs as i64) + slot * CE_NE * 8
132 let vp: *i64 = vp_i as *i64
133 let t0: i64 = sys_now_ms()
134 let n: i64 = nsv_embed(txt, l, 1, vp)
135 let t1: i64 = sys_now_ms()
136 ce_ws("[embed] " as *u8)
137 ce_kv("slot" as *u8, slot)
138 ce_kv("ntok" as *u8, n)
139 ce_kv("ms" as *u8, t1 - t0)
140 ce_ws("\n" as *u8)
141 return n
142}
143
144func ce_vp(vecs: *i64, slot: i64) -> *i64 {
145 let p: i64 = (vecs as i64) + slot * CE_NE * 8
146 return p as *i64
147}
148
149func main() -> i64 {
150 ce_ws("[embed-gate] sovereign code embedding (jina recipe arXiv:2508.21290) on no-float coder, mode=i8\n" as *u8)
151 let ir: i64 = nsv_init("/home/elderwesto/nx_stage/nx_coder_model.gguf\x00" as *u8)
152 if ir != 0 {
153 ce_kv("[embed-gate] INIT-FAIL rc" as *u8, ir)
154 ce_ws("\n" as *u8)
155 return 1
156 }
157 let vecs: *i64 = sys_mmap(CE_SLOTS * CE_NE * 8) as *i64
158 // jina EXACT task instruction prefixes (nl2code)
159 let qp: *u8 = "Find the most relevant code snippet given the following query:\n" as *u8
160 let pp: *u8 = "Candidate code snippet:\n" as *u8
161 // jina model-card 2x2 demo (bf16 jina-tuned oracle: q1p1=817 q1p2=124 q2p1=120 q2p2=553 pm)
162 var rc: i64 = 0
163 rc = ce_embed(vecs, 0, qp, "print hello world in python" as *u8)
164 if rc < 1 { return 1 }
165 rc = ce_embed(vecs, 1, qp, "initialize array of 5 zeros in c++" as *u8)
166 if rc < 1 { return 1 }
167 rc = ce_embed(vecs, 2, pp, "print('Hello World!')" as *u8)
168 if rc < 1 { return 1 }
169 rc = ce_embed(vecs, 3, pp, "int arr[5] = {0, 0, 0, 0, 0};" as *u8)
170 if rc < 1 { return 1 }
171 // Nishi-corpus 3x3 micro-retrieval (REAL API shapes from runtime/)
172 rc = ce_embed(vecs, 4, qp, "read a u32 metadata value from a gguf header by key" as *u8)
173 if rc < 1 { return 1 }
174 rc = ce_embed(vecs, 5, qp, "connect to a unix domain socket at a filesystem path" as *u8)
175 if rc < 1 { return 1 }
176 rc = ce_embed(vecs, 6, qp, "rms normalize a hidden row with gamma weights in fixed point" as *u8)
177 if rc < 1 { return 1 }
178 rc = ce_embed(vecs, 7, pp, "nx_gguf_meta_find(buf, flen, hdr, key, klen, voff, vty)\nlet v: i64 = nx_gguf_meta_read_u32(buf, voff[0])" as *u8)
179 if rc < 1 { return 1 }
180 rc = ce_embed(vecs, 8, pp, "let fd: i64 = sys_socket(1, 1, 0)\nsa[0] = 1 as u8\nsys_connect(fd, sa as *u8, 110)" as *u8)
181 if rc < 1 { return 1 }
182 rc = ce_embed(vecs, 9, pp, "rmsnorm_gamma_row_q24(h1, gout, 0, ne, normed, 0)" as *u8)
183 if rc < 1 { return 1 }
184 // B: re-embed q1 LAST (after every other text) -> must be BITWISE identical to slot 0
185 rc = ce_embed(vecs, 10, qp, "print hello world in python" as *u8)
186 if rc < 1 { return 1 }
187 ce_kv("[scale] maxabs_q24_slot0" as *u8, ce_maxabs(ce_vp(vecs, 0)))
188 ce_ws("\n" as *u8)
189 // ---- jina 2x2 ----
190 let c11: i64 = ce_cos_pm(ce_vp(vecs, 0), ce_vp(vecs, 2))
191 let c12: i64 = ce_cos_pm(ce_vp(vecs, 0), ce_vp(vecs, 3))
192 let c21: i64 = ce_cos_pm(ce_vp(vecs, 1), ce_vp(vecs, 2))
193 let c22: i64 = ce_cos_pm(ce_vp(vecs, 1), ce_vp(vecs, 3))
194 ce_ws("[jina-2x2 pm] " as *u8)
195 ce_kv("q1p1" as *u8, c11)
196 ce_kv("q1p2" as *u8, c12)
197 ce_kv("q2p1" as *u8, c21)
198 ce_kv("q2p2" as *u8, c22)
199 ce_ws("(jina bf16 oracle: 817 124 120 553)\n" as *u8)
200 var gate_a: i64 = 0
201 if c11 > c12 { if c22 > c21 { gate_a = 1 } }
202 // ---- determinism / state-leak ----
203 let gate_b: i64 = ce_bytes_eq(ce_vp(vecs, 0), ce_vp(vecs, 10))
204 let selfpm: i64 = ce_cos_pm(ce_vp(vecs, 0), ce_vp(vecs, 10))
205 var gate_c: i64 = 0
206 if selfpm >= 999 { gate_c = 1 }
207 ce_ws("[determinism] " as *u8)
208 ce_kv("bitwise_eq" as *u8, gate_b)
209 ce_kv("self_pm" as *u8, selfpm)
210 ce_ws("\n" as *u8)
211 // ---- nishi 3x3 retrieval ----
212 var hits: i64 = 0
213 var r: i64 = 0
214 while r < 3 {
215 let s0: i64 = ce_cos_pm(ce_vp(vecs, 4 + r), ce_vp(vecs, 7))
216 let s1: i64 = ce_cos_pm(ce_vp(vecs, 4 + r), ce_vp(vecs, 8))
217 let s2: i64 = ce_cos_pm(ce_vp(vecs, 4 + r), ce_vp(vecs, 9))
218 var am: i64 = 0
219 var best: i64 = s0
220 if s1 > best { am = 1; best = s1 }
221 if s2 > best { am = 2; best = s2 }
222 ce_ws("[nishi-3x3] " as *u8)
223 ce_kv("row" as *u8, r)
224 ce_kv("s0" as *u8, s0)
225 ce_kv("s1" as *u8, s1)
226 ce_kv("s2" as *u8, s2)
227 ce_kv("argmax" as *u8, am)
228 if am == r { hits = hits + 1; ce_ws("HIT\n" as *u8) } else { ce_ws("MISS\n" as *u8) }
229 r = r + 1
230 }
231 ce_kv("[retrieval] hits" as *u8, hits)
232 ce_ws("/3\n" as *u8)
233 // ---- E: the API route (synthetic HTTP request through nsv_handle; the daemon serves these bytes) ----
234 let req: *u8 = sys_mmap(4096)
235 let qn: i64 = nsv_cat(req, 0, "POST /embed HTTP/1.1\r\nContent-Type: application/json\r\n\r\n{\"text\":\"open a file and read all of its bytes\",\"task\":\"nl2code\",\"kind\":\"query\",\"mode\":\"i8\"}" as *u8)
236 let resb: *u8 = sys_mmap(65536)
237 let rl: i64 = nsv_handle(req, qn, resb, 65536)
238 var gate_e: i64 = 0
239 if rl > 60 { if resb[9]==(50 as u8) { if resb[10]==(48 as u8) { if resb[11]==(48 as u8) {
240 if ce_find(resb, rl, "\"ok\":1" as *u8) >= 0 { if ce_find(resb, rl, "\"dim\":896" as *u8) >= 0 { gate_e = 1 } }
241 } } } }
242 ce_ws("[api] POST /embed via nsv_handle " as *u8)
243 ce_kv("resp_len" as *u8, rl)
244 ce_kv("gate_e" as *u8, gate_e)
245 ce_ws("\n[api] resp head: " as *u8)
246 var eh: i64 = rl
247 if eh > 200 { eh = 200 }
248 sys_write(1, resb, eh)
249 ce_ws("...\n" as *u8)
250 // ---- verdict ----
251 var green: i64 = 0
252 if gate_a == 1 { if gate_b == 1 { if gate_c == 1 { if hits == 3 { if gate_e == 1 { green = 1 } } } } }
253 if green == 1 {
254 ce_ws("VERDICT GREEN sovereign-code-embed cpu-i8 (A order + B bitwise-determinism + C self + D 3/3 + E api-route; pipeline correctness on BASE weights -- SOTA grading pends jina-weight oracle)\n" as *u8)
255 return 0
256 }
257 ce_ws("VERDICT RED " as *u8)
258 ce_kv("A" as *u8, gate_a)
259 ce_kv("B" as *u8, gate_b)
260 ce_kv("C" as *u8, gate_c)
261 ce_kv("D" as *u8, hits)
262 ce_kv("E" as *u8, gate_e)
263 ce_ws("\n" as *u8)
264 return 1
265}