nx_nl_concept_miner_gate.nx source

1import "nx_gate_gn.nx" 2import "nx_gate_base.nx" 3// nx_nl_concept_miner_gate.nx -- ATTACK THE RESIDUAL: free-form NATURAL-LANGUAGE translation (the one piece the 4// structured parser nx_primitive_miner did not cover). A sovereign NL concept-miner recognizes math concepts 5// described in PROSE ("the sum of the cubes" -> exponent 3) via a data-driven phrase->concept table, NO LLM, 6// then runs select+verify. HONEST: this is SHALLOW semantic extraction -- it only knows the concepts in its 7// table (it cannot understand a TRULY novel description; that head-start is the LLM's broad pretraining). BUT 8// the table GROWS from experience (wake-sleep), so the residual is a COVERAGE gap that CLOSES, not a capability 9// wall. The honest claim: the LLM's NL edge = coverage/cold-start, replaceable by an experience-grown table. 10// T1 NL-EXTRACT: from PROSE (concept words, not i^k tokens) the miner extracts {S3,S4}; irrelevant prose -> 0. 11// T2 SELECT: of the prose-mined {S3,S4}, only S3 makes x^4 solvable+generalizing -> kept. 12// T3 END-TO-END: prose -> NL-mine -> select -> the expanded grammar solves x^4 (held-out 625/1296), NO LLM. 13// T4 THE TRUE RESIDUAL: prose describing a concept OUTSIDE the table ("hyperfactorial") -> 0 -> honest escalate. 14// T5 WAKE-SLEEP: a phrasing the table lacks ("fourth powers") -> unparseable -> LEARN it -> now parseable -> solves. 15// expect_exit: 0 license_tier: ORIGINAL 16import "nx_synth_fit.nx" 17import "nx_syscalls.nx" 18 19func grow(name: *u8, ok: i64) -> i64 { if ok==1 { gw(" PASS " as *u8) } else { gw(" FAIL " as *u8) } gw(name); gw(" 20" as *u8); return ok } 21func slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} return n } 22// substring search: 1 if needle occurs in hay. 23func contains(hay: *u8, hlen: i64, needle: *u8, nlen: i64) -> i64 { 24 var i: i64=0 25 while (i+nlen)<=hlen { 26 var j: i64=0; var eq: i64=1 27 while j<nlen { let a: *u8=((hay as i64)+i+j) as *u8; if a[0]!=needle[j] { eq=0; j=nlen } else { j=j+1 } } 28 if eq==1 { return 1 } 29 i=i+1 30 } 31 return 0 32} 33 34// the concept table (the "ontology", grown from experience): phrase -> power-sum exponent. tn = #entries. 35// requires a "sum" trigger in the text (a power-SUM). NO LLM -- shallow phrase recognition. 36func nl_mine(text: *u8, tphr: *i64, texp: *i64, tn: i64, out: *i64) -> i64 { 37 let tl: i64=slen(text) 38 if contains(text,tl,"sum" as *u8,3)==0 { return 0 } // power-sum trigger absent -> no candidate 39 var cnt: i64=0; var e: i64=0 40 while e<tn { 41 let ph: *u8=tphr[e] as *u8 42 if contains(text,tl,ph,slen(ph))==1 { 43 let k: i64=texp[e] 44 var found: i64=0; var j: i64=0; while j<cnt { if out[j]==k { found=1 } j=j+1 } 45 if found==0 { out[cnt]=k; cnt=cnt+1 } 46 } 47 e=e+1 48 } 49 return cnt 50} 51 52func main() -> i64 { 53 gw("=== nx_nl_concept_miner: free-form NL -> concept -> primitive (attack the residual), no LLM ===\n" as *u8) 54 var pass: i64=0; var total: i64=0 55 56 // the concept table (phrase -> exponent). Built from experience; grows in T5. 57 let tphr: *i64=sys_mmap(128) as *i64; let texp: *i64=sys_mmap(128) as *i64 58 tphr[0]="squares" as *u8 as i64; texp[0]=2 59 tphr[1]="cubes" as *u8 as i64; texp[1]=3 60 tphr[2]="fourth powers" as *u8 as i64; texp[2]=4 61 tphr[3]="fifth powers" as *u8 as i64; texp[3]=5 62 var tn: i64=4 63 64 // TARGET A = x^4 (out of base grammar). train x=0..4, held {5->625, 6->1296}. 65 let xs: *i64=sys_mmap(64) as *i64; let ys: *i64=sys_mmap(64) as *i64 66 xs[0]=0; xs[1]=1; xs[2]=2; xs[3]=3; xs[4]=4 67 ys[0]=0; ys[1]=1; ys[2]=16; ys[3]=81; ys[4]=256 68 let n: i64=5 69 70 // T1: NL-EXTRACT from PROSE (words, not i^k). doc has "cubes" and "fourth powers" + "sum". 71 let doc: *u8="The closed form likely involves the sum of the cubes of integers; the sum of fourth powers is a tempting dead end." as *u8 72 let exps: *i64=sys_mmap(128) as *i64 73 let cnt: i64=nl_mine(doc,tphr,texp,tn,exps) 74 let irrel: *u8="The weather is nice today and the market index rose three percent." as *u8 75 let iexps: *i64=sys_mmap(64) as *i64 76 let icnt: i64=nl_mine(irrel,tphr,texp,tn,iexps) 77 var got34: i64=0; if cnt==2 { if exps[0]==3 { if exps[1]==4 { got34=1 } } } 78 total=total+1; if got34==1 { if icnt==0 { pass=pass+1; gw(" [PASS] " as *u8) } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } 79 gw("T1 NL-EXTRACT: from PROSE the miner extracted " as *u8); gn(cnt); gw(" concepts {" as *u8); var p: i64=0; while p<cnt { gw("S" as *u8); gn(exps[p]); if p<cnt-1 { gw("," as *u8) } p=p+1 } gw("} (concept WORDS, not i^k), and " as *u8); gn(icnt); gw(" from irrelevant prose\n" as *u8) 80 81 // T2: SELECT+VERIFY over the prose-mined candidates -- only S3 kept. 82 var selected: i64=0; var nkept: i64=0 83 var ci: i64=0 84 while ci<cnt { 85 let k: i64=exps[ci] 86 let keep: i64=sf_keep(k,xs,ys,n,5,625,6,1296) 87 gw(" prose concept S" as *u8); gn(k); gw(": kept=" as *u8); gn(keep); gw("\n" as *u8) 88 if keep==1 { selected=k; nkept=nkept+1 } 89 ci=ci+1 90 } 91 total=total+1; if selected==3 { if nkept==1 { pass=pass+1; gw(" [PASS] " as *u8) } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } 92 gw("T2 SELECT+VERIFY: of prose-mined {S3,S4} kept exactly " as *u8); gn(nkept); gw(" -> selected S" as *u8); gn(selected); gw(", no LLM\n" as *u8) 93 94 // T3: END-TO-END prose -> NL-mine -> select -> solve x^4. 95 let bb: *i64=sys_mmap(64) as *i64 96 let okfit: i64=sf_fit(3,xs,ys,n,12345,bb) 97 let h5: i64=sf_model(bb,3,5); let h6: i64=sf_model(bb,3,6) 98 total=total+1; if okfit==1 { if h5==625 { if h6==1296 { pass=pass+1; gw(" [PASS] " as *u8) } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } 99 gw("T3 END-TO-END: prose->NL-mine->select-> base+S3 solves x^4 held-out x=5->" as *u8); gn(h5); gw(" x=6->" as *u8); gn(h6); gw("\n" as *u8) 100 101 // T4: THE TRUE RESIDUAL -- a concept OUTSIDE the table (prose mentions "hyperfactorial") -> 0 -> escalate. 102 let unk: *u8="To solve this you likely need the sum of the hyperfactorial of n, an exotic construct." as *u8 103 let uexps: *i64=sys_mmap(64) as *i64 104 let ucnt: i64=nl_mine(unk,tphr,texp,tn,uexps) 105 total=total+1; if ucnt==0 { pass=pass+1; gw(" [PASS] " as *u8) } else { gw(" [FAIL] " as *u8) } 106 gw("T4 TRUE RESIDUAL: prose with a concept OUTSIDE the table (hyperfactorial) -> " as *u8); gn(ucnt); gw(" candidates -> honest escalate (THIS is the LLM's remaining edge: cold-start coverage)\n" as *u8) 107 108 // T5: WAKE-SLEEP -- a phrasing the table LACKS, then LEARNED. Target B = sum of fourth powers (powsum4), degree 5. 109 let xs4: *i64=sys_mmap(64) as *i64; let ys4: *i64=sys_mmap(64) as *i64 110 var z: i64=0; while z<6 { xs4[z]=z; ys4[z]=sf_powsum(4,z); z=z+1 } // 0,1,17,98,354,979 111 let doc4: *u8="This quantity is the sum of fourth powers of the first n integers." as *u8 112 // round 1: table WITHOUT "fourth powers" (only squares,cubes) -> unparseable. 113 let t1phr: *i64=sys_mmap(64) as *i64; let t1exp: *i64=sys_mmap(64) as *i64 114 t1phr[0]="squares" as *u8 as i64; t1exp[0]=2; t1phr[1]="cubes" as *u8 as i64; t1exp[1]=3 115 let r1: *i64=sys_mmap(64) as *i64 116 let c1: i64=nl_mine(doc4,t1phr,t1exp,2,r1) 117 // LEARN the phrasing (wake-sleep): add "fourth powers"->4. 118 t1phr[2]="fourth powers" as *u8 as i64; t1exp[2]=4 119 let r2: *i64=sys_mmap(64) as *i64 120 let c2: i64=nl_mine(doc4,t1phr,t1exp,3,r2) 121 var solved4: i64=0 122 if c2==1 { if r2[0]==4 { solved4=sf_keep(4,xs4,ys4,6,6,sf_powsum(4,6),7,sf_powsum(4,7)) } } 123 total=total+1; if c1==0 { if c2==1 { if solved4==1 { pass=pass+1; gw(" [PASS] " as *u8) } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } } else { gw(" [FAIL] " as *u8) } 124 gw("T5 WAKE-SLEEP: 'fourth powers' unknown -> " as *u8); gn(c1); gw(" candidates (escalate); LEARN the phrasing -> " as *u8); gn(c2); gw(" candidate -> sum-of-4th-powers SOLVED=" as *u8); gn(solved4); gw(" = coverage GREW from experience, no LLM\n" as *u8) 125 126 gw("\n THE RESIDUAL, HONESTLY: the LLM's last edge = free-form NL translation. A sovereign concept-miner covers it for\n" as *u8) 127 gw(" KNOWN concept families (recognize the concept in prose -> primitive), and the table GROWS from experience (T5 wake-sleep) --\n" as *u8) 128 gw(" so the gap is COVERAGE/cold-start (T4), not a capability wall. The LLM starts with internet-scale coverage; we grow ours.\n" as *u8) 129 gw(" Truly novel concepts before they're learned are where the LLM head-starts -- the same cold-start frontier as the learned prior.\n" as *u8) 130 gw("NL-CONCEPT-MINER verdict=" as *u8) 131 if pass==total { gw("GREEN passes=" as *u8); gn(pass); gw("/" as *u8); gn(total); gw(" -- free-form NL translation covered for known/learned concepts, grows from experience, NO LLM\n" as *u8); sys_exit(0); return 0 } 132 gw("RED passes=" as *u8); gn(pass); gw("/" as *u8); gn(total); gw("\n" as *u8); sys_exit(1); return 1 133}