spear-kernels 1.7.0 → 1.9.0

package/README.md

@@ -42,5 +42,5 @@ instance.exports.spear(2); // same protected-division semantics as the JS backen
 ```
 
 Division is **protected** exactly like the engine: denominators floored at
-±1e-4 (sign-preserving), results clamped to ±1e4. Discovered formulas rely on
-these rails — do not swap in bare `/`.
+±1e-4 (sign-preserving), results clamped to ±1e4. Discovered formulas rely on
+these rails — do not swap in bare `/`.

package/index.cjs

@@ -1,4 +1,4 @@
-// Auto-generated by scripts/gen-kernels-package.ts — DO NOT EDIT.
+// Auto-generated by scripts/gen-kernels-package.ts — DO NOT EDIT.
 'use strict';
 const kernels = {
   "silu": {
@@ -29,11 +29,11 @@ const kernels = {
   "gelu": {
     id: "gelu",
     precise:     {
-      js: "((x) => ((0.999993 * ((0.5 * x) * (1 + Math.tanh((0.797885 * (x + (0.044715 * (x) * (x) * (x)))))))) + -0.000058))",
-      eval: ((x) => ((0.999993 * ((0.5 * x) * (1 + Math.tanh((0.797885 * (x + (0.044715 * (x) * (x) * (x)))))))) + -0.000058)),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.999993f * ((0.5f * x) * (1f + tanhf((0.797885f * (x + (0.044715f * powf(x, 3.0f)))))))) + -0.000058f);\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.999993 * ((0.5 * x) * (1 + torch.tanh((0.797885 * (x + (0.044715 * (x ** 3)))))))) + -0.000058)",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgwBA2VudgR0YW5oAAEDAgEABwkBBXNwZWFyAAEKUAFOAERhU+dR8f/vP0QAAAAAAADgPyAAokQAAAAAAADwP0S4AZ8fRojpPyAARPcBSG3i5KY/IAAgACAAoqKioKIQAKCiokSQvknToGgOv6AL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return ((0.5 * x) * (1 + __erf((x * 0.7071067811865475)))); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return ((0.5 * x) * (1 + __erf((x * 0.7071067811865475)))); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.5f * x) * (1f + erff((x * 0.707107f))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.5 * x) * (1 + torch.erf((x * 0.707107))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQopAScARAAAAAAAAOA/IACiRAAAAAAAAPA/IABEzDt/Zp6g5j+iEACgogs=",
     },
     fast:     {
       js: "((x) => Math.max(0, x))",
@@ -43,11 +43,11 @@ const kernels = {
       wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAAChABDgAgAEQAAAAAAAAAAKUL",
     },
     meta: {
-      metric: 3.55135869310835e-8,
+      metric: 1.5455587803707568e-34,
       level: 5,
-      formulaCost: 11,
+      formulaCost: 5,
       exactCost: 46,
-      speedupVsExact: 4.181818181818182,
+      speedupVsExact: 9.2,
       vsIterative: null,
     },
   },
@@ -919,25 +919,25 @@ const kernels = {
   "erf_prob": {
     id: "erf_prob",
     precise:     {
-      js: "((x) => (1.106774 * Math.min(1e4, Math.max(-1e4, ((x + (0.034298 * (x) * (x) * (x)))) / (((0.995 + (0.378089 * (x) * (x)))) < 1e-4 && ((0.995 + (0.378089 * (x) * (x)))) > -1e-4 ? (((0.995 + (0.378089 * (x) * (x)))) >= 0 ? 1e-4 : -1e-4) : ((0.995 + (0.378089 * (x) * (x)))))))))",
-      eval: ((x) => (1.106774 * Math.min(1e4, Math.max(-1e4, ((x + (0.034298 * (x) * (x) * (x)))) / (((0.995 + (0.378089 * (x) * (x)))) < 1e-4 && ((0.995 + (0.378089 * (x) * (x)))) > -1e-4 ? (((0.995 + (0.378089 * (x) * (x)))) >= 0 ? 1e-4 : -1e-4) : ((0.995 + (0.378089 * (x) * (x))))))))),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (1.106774f * fminf(fmaxf(((x + (0.034298f * powf(x, 3.0f))) / copysignf(fmaxf(fabsf((0.995f + (0.378089f * powf(x, 2.0f)))), 1.0e-4f), (0.995f + (0.378089f * powf(x, 2.0f))))), -1.0e4f), 1.0e4f));\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (1.106774 * ((x + (0.034298 * (x ** 3))) / (abs((0.995 + (0.378089 * (x * x)))) < 1e-4 ? 1e-4 : (0.995 + (0.378089 * (x * x))))))",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACnYBdABEeQJhp1i18T8gAEQp7KLogY+hPyAAIAAgAKKioqBE16NwPQrX7z9Ezox+NJwy2D8gACAAoqKgmUQtQxzr4jYaP6VE16NwPQrX7z9Ezox+NJwy2D8gACAAoqKgpqNEAAAAAACIw8ClRAAAAAAAiMNApKIL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return erff(x);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.erf(x)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQoIAQYAIAAQAAs=",
     },
     fast:     {
-      js: "((x) => Math.atan(x))",
-      eval: ((x) => Math.atan(x)),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return atanf(x);\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.atan(x)",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgwBA2VudgRhdGFuAAEDAgEABwkBBXNwZWFyAAEKCAEGACAAEAAL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return erff(x);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.erf(x)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQoIAQYAIAAQAAs=",
     },
     meta: {
-      metric: 0.00004650870031371604,
-      level: 3,
-      formulaCost: 12,
+      metric: 0,
+      level: 5,
+      formulaCost: 1,
       exactCost: 26,
-      speedupVsExact: 2.1666666666666665,
+      speedupVsExact: 26,
       vsIterative: null,
     },
   },
@@ -2002,8 +2002,158 @@ const kernels = {
       vsIterative: null,
     },
   },
+  "rope_freq": {
+    id: "rope_freq",
+    precise:     {
+      js: "((x) => ((0.9999991529919943 * ((0.999947 * Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))) + -0.000038)) + 0.00001879512303862185))",
+      eval: ((x) => ((0.9999991529919943 * ((0.999947 * Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))) + -0.000038)) + 0.00001879512303862185)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.999999f * ((0.999947f * cosf((1023.994f * expf(fmaxf(-50.0f, fminf(50.0f, (-0.28782f * x))))))) + -0.000038f)) + 0.000019f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.999999 * ((0.999947 * torch.cos((1023.994 * torch.exp(torch.clamp((-0.28782 * x), -50.0, 50.0))))) + -0.000038)) + 0.000019)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AhUCA2VudgNleHAAAQNlbnYDY29zAAEDAgEABwkBBXNwZWFyAAIKWgFYAEQqG0Q5/v/vP0RLd9fZkP/vP0TLoUW28/+PQERBn8iTpGvSvyAAokQAAAAAAABJwKVEAAAAAAAASUCkEACiEAGiRBgK2A5G7AO/oKJEuB45BUe18z6gCw==",
+    },
+    fast:     {
+      js: "((x) => Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x)))))))",
+      eval: ((x) => Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return cosf((1023.994f * expf(fmaxf(-50.0f, fminf(50.0f, (-0.28782f * x))))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.cos((1023.994 * torch.exp(torch.clamp((-0.28782 * x), -50.0, 50.0))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AhUCA2VudgNleHAAAQNlbnYDY29zAAEDAgEABwkBBXNwZWFyAAIKMgEwAETLoUW28/+PQERBn8iTpGvSvyAAokQAAAAAAABJwKVEAAAAAAAASUCkEACiEAEL",
+    },
+    meta: {
+      metric: 8.450097215964813e-7,
+      level: 5,
+      formulaCost: 8,
+      exactCost: 43,
+      speedupVsExact: 5.375,
+      vsIterative: null,
+    },
+  },
+  "bilateral_weight": {
+    id: "bilateral_weight",
+    precise:     {
+      js: "((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2)))))))",
+      eval: ((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return fminf(fmaxf((expf(fmaxf(-50.0f, fminf(50.0f, (-(powf(x, 2.0f) * 50f))))) / copysignf(fmaxf(fabsf((1f + (powf(x, 2.0f) * 2f))), 1.0e-4f), (1f + (powf(x, 2.0f) * 2f)))), -1.0e4f), 1.0e4f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (torch.exp(torch.clamp((-((x * x) * 50)), -50.0, 50.0)) / (abs((1 + ((x * x) * 2))) < 1e-4 ? 1e-4 : (1 + ((x * x) * 2))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNleHAAAQMCAQAHCQEFc3BlYXIAAQp9AXsAIAAgAKJEAAAAAAAASUCimkQAAAAAAABJwKVEAAAAAAAASUCkEABEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgmUQtQxzr4jYaP6VEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgpqNEAAAAAACIw8ClRAAAAAAAiMNApAs=",
+    },
+    fast:     {
+      js: "((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2)))))))",
+      eval: ((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return fminf(fmaxf((expf(fmaxf(-50.0f, fminf(50.0f, (-(powf(x, 2.0f) * 50f))))) / copysignf(fmaxf(fabsf((1f + (powf(x, 2.0f) * 2f))), 1.0e-4f), (1f + (powf(x, 2.0f) * 2f)))), -1.0e4f), 1.0e4f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (torch.exp(torch.clamp((-((x * x) * 50)), -50.0, 50.0)) / (abs((1 + ((x * x) * 2))) < 1e-4 ? 1e-4 : (1 + ((x * x) * 2))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNleHAAAQMCAQAHCQEFc3BlYXIAAQp9AXsAIAAgAKJEAAAAAAAASUCimkQAAAAAAABJwKVEAAAAAAAASUCkEABEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgmUQtQxzr4jYaP6VEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgpqNEAAAAAACIw8ClRAAAAAAAiMNApAs=",
+    },
+    meta: {
+      metric: 9.336399957342636e-35,
+      level: 5,
+      formulaCost: 11,
+      exactCost: 45,
+      speedupVsExact: 4.090909090909091,
+      vsIterative: null,
+    },
+  },
+  "bs_d1_sigma": {
+    id: "bs_d1_sigma",
+    precise:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((Math.abs(x)) < 1e-4 && (Math.abs(x)) > -1e-4 ? ((Math.abs(x)) >= 0 ? 1e-4 : -1e-4) : (Math.abs(x))))) + (0.25 * Math.abs(x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((Math.abs(x)) < 1e-4 && (Math.abs(x)) > -1e-4 ? ((Math.abs(x)) >= 0 ? 1e-4 : -1e-4) : (Math.abs(x))))) + (0.25 * Math.abs(x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(fabsf(x)), 1.0e-4f), fabsf(x))), -1.0e4f), 1.0e4f) + (0.25f * fabsf(x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(torch.abs(x)) < 1e-4 ? 1e-4 : torch.abs(x))) + (0.25 * torch.abs(x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkIBQABEop9GrXKZyz8gAJmZRC1DHOviNho/pSAAmaajRAAAAAAAiMPApUQAAAAAAIjDQKREAAAAAAAA0D8gAJmioAs=",
+    },
+    fast:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (3.635301) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (4.215319 * x)))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (3.635301) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (4.215319 * x))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((3.635301f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (4.215319f * x));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((3.635301 / (abs(x) < 1e-4 ? 1e-4 : x)) + (4.215319 * x))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACj8BPQBExO3QsBgVDUAgAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRGfUfJV83BBAIACioAs=",
+    },
+    meta: {
+      metric: 1.6217711315014304e-18,
+      level: 5,
+      formulaCost: 8,
+      exactCost: 9,
+      speedupVsExact: 1.125,
+      vsIterative: null,
+    },
+  },
+  "mel_scale": {
+    id: "mel_scale",
+    precise:     {
+      js: "((x) => (1126.994181 * Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286))))))",
+      eval: ((x) => (1126.994181 * Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286)))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (1126.994181f * logf(fmaxf(1.0e-30f, (1f + (x * 0.001429f)))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (1126.994181 * torch.log(torch.clamp((1 + (x * 0.001429)), min=1e-30)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQowAS4ARDiFlQr6m5FARAAAAAAAAPA/IABEsak05NxnVz+ioESgwuv+S0i0OaUQAKIL",
+    },
+    fast:     {
+      js: "((x) => Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286)))))",
+      eval: ((x) => Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return logf(fmaxf(1.0e-30f, (1f + (x * 0.001429f))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.log(torch.clamp((1 + (x * 0.001429)), min=1e-30))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQomASQARAAAAAAAAPA/IABEsak05NxnVz+ioESgwuv+S0i0OaUQAAs=",
+    },
+    meta: {
+      metric: 1.4594066099947563e-12,
+      level: 5,
+      formulaCost: 4,
+      exactCost: 23,
+      speedupVsExact: 5.75,
+      vsIterative: null,
+    },
+  },
+  "bs_d2_sigma": {
+    id: "bs_d2_sigma",
+    precise:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (-(0.25f * x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(x) < 1e-4 ? 1e-4 : x)) + (-(0.25 * x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkABPgBEop9GrXKZyz8gAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRAAAAAAAANA/IACimqAL",
+    },
+    fast:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (-(0.25f * x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(x) < 1e-4 ? 1e-4 : x)) + (-(0.25 * x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkABPgBEop9GrXKZyz8gAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRAAAAAAAANA/IACimqAL",
+    },
+    meta: {
+      metric: 1.6217711130493227e-18,
+      level: 5,
+      formulaCost: 7,
+      exactCost: 11,
+      speedupVsExact: 1.5714285714285714,
+      vsIterative: null,
+    },
+  },
+  "a_weighting": {
+    id: "a_weighting",
+    precise:     {
+      js: "((x) => ((0.9803615530479888 * ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)) + 0.039276893904016105))",
+      eval: ((x) => ((0.9803615530479888 * ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)) + 0.039276893904016105)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.980362f * ((8.859884f * logf(fmaxf(1.0e-30f, fminf(fmaxf((((148691236f * powf(x, 2.0f)) * powf(x, 2.0f)) / copysignf(fmaxf(fabsf(((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f)))), 1.0e-4f), ((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f))))), -1.0e4f), 1.0e4f)))) + 2f)) + 0.039277f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.980362 * ((8.859884 * torch.log(torch.clamp((((148691236 * (x * x)) * (x * x)) / (abs((((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236)))) < 1e-4 ? 1e-4 : (((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236))))), min=1e-30))) + 2)) + 0.039277)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQrxAQHuAQBEphkTMR9f7z9EE7U0t0K4IUBEAAAASLK5oUEgACAAoqIgACAAoqIgACAAokT2KFyPwoV6QKAgACAAokTsUbgepafGQKAgACAAokQfhevR4J0gQaCimZ8gACAAokQAAABIsrmhQaCioplELUMc6+I2Gj+lIAAgAKJE9ihcj8KFekCgIAAgAKJE7FG4HqWnxkCgIAAgAKJEH4Xr0eCdIEGgopmfIAAgAKJEAAAASLK5oUGgoqKmo0QAAAAAAIjDwKVEAAAAAACIw0CkRKDC6/5LSLQ5pRAAokQAAAAAAAAAQKCiRK7HnN0ZHKQ/oAs=",
+    },
+    fast:     {
+      js: "((x) => ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2))",
+      eval: ((x) => ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((8.859884f * logf(fmaxf(1.0e-30f, fminf(fmaxf((((148691236f * powf(x, 2.0f)) * powf(x, 2.0f)) / copysignf(fmaxf(fabsf(((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f)))), 1.0e-4f), ((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f))))), -1.0e4f), 1.0e4f)))) + 2f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((8.859884 * torch.log(torch.clamp((((148691236 * (x * x)) * (x * x)) / (abs((((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236)))) < 1e-4 ? 1e-4 : (((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236))))), min=1e-30))) + 2)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQrdAQHaAQBEE7U0t0K4IUBEAAAASLK5oUEgACAAoqIgACAAoqIgACAAokT2KFyPwoV6QKAgACAAokTsUbgepafGQKAgACAAokQfhevR4J0gQaCimZ8gACAAokQAAABIsrmhQaCioplELUMc6+I2Gj+lIAAgAKJE9ihcj8KFekCgIAAgAKJE7FG4HqWnxkCgIAAgAKJEH4Xr0eCdIEGgopmfIAAgAKJEAAAASLK5oUGgoqKmo0QAAAAAAIjDwKVEAAAAAACIw0CkRKDC6/5LSLQ5pRAAokQAAAAAAAAAQKAL",
+    },
+    meta: {
+      metric: 2.7800484820648814e-29,
+      level: 5,
+      formulaCost: 26,
+      exactCost: 55,
+      speedupVsExact: 2.1153846153846154,
+      vsIterative: null,
+    },
+  },
 };
 
-const kernelIds = ["silu","gelu","sigmoid","kv_cache","free_fall","kepler","gaussian_cdf","damped_pendulum","rl_distillation","lambert_w","rc_circuit","gaussian_kernel","diffusion_beta","bilinear_interp","european_call","temporal_grad","lorentz","hill","kerr","lennard_jones","damped_oscillation","logistic_growth","softplus","kdv_soliton","kerr_spin","pendulum_hybrid","eigen3_sym","ik_reach","idm_following","gemv4","rope_rot","gauss_shader","ema_smooth","smoothstep","srgb_gamma","tanh_sat","atan_unit","srgb_decode","erf_prob","huber_loss","cosh_curve","bessel_j0","bessel_j1","blackbody_r","aces_fit","logsumexp2","probit_quantile","pmt_finance","blackbody_g","blackbody_b","bessel_j2","logit_ml","kelly_criterion","rsi_momentum","implied_vol","michaelis_menten","temperature_softmax","doppler_effect","stefan_boltzmann","mm1_queue_wait","grover_amplitude","concurrence_pure","chsh_correlation","legendre_p2","laguerre_l2","asin_hard","qfi_dephasing","amp_damp_fid","loschmidt_rate","bessel_i0e","elliptic_k","kepler_solver","fast_exp_alu","fresnel_schlick","rayleigh_phase","smootherstep","fog_exp2","back_ease_out","light_falloff_punctual","bias_slope","uncharted2_tonemap","mish"];
+const kernelIds = ["silu","gelu","sigmoid","kv_cache","free_fall","kepler","gaussian_cdf","damped_pendulum","rl_distillation","lambert_w","rc_circuit","gaussian_kernel","diffusion_beta","bilinear_interp","european_call","temporal_grad","lorentz","hill","kerr","lennard_jones","damped_oscillation","logistic_growth","softplus","kdv_soliton","kerr_spin","pendulum_hybrid","eigen3_sym","ik_reach","idm_following","gemv4","rope_rot","gauss_shader","ema_smooth","smoothstep","srgb_gamma","tanh_sat","atan_unit","srgb_decode","erf_prob","huber_loss","cosh_curve","bessel_j0","bessel_j1","blackbody_r","aces_fit","logsumexp2","probit_quantile","pmt_finance","blackbody_g","blackbody_b","bessel_j2","logit_ml","kelly_criterion","rsi_momentum","implied_vol","michaelis_menten","temperature_softmax","doppler_effect","stefan_boltzmann","mm1_queue_wait","grover_amplitude","concurrence_pure","chsh_correlation","legendre_p2","laguerre_l2","asin_hard","qfi_dephasing","amp_damp_fid","loschmidt_rate","bessel_i0e","elliptic_k","kepler_solver","fast_exp_alu","fresnel_schlick","rayleigh_phase","smootherstep","fog_exp2","back_ease_out","light_falloff_punctual","bias_slope","uncharted2_tonemap","mish","rope_freq","bilateral_weight","bs_d1_sigma","mel_scale","bs_d2_sigma","a_weighting"];
 
 module.exports = { kernels, kernelIds, default: kernels };

package/index.js

@@ -1,4 +1,4 @@
-// Auto-generated by scripts/gen-kernels-package.ts — DO NOT EDIT.
+// Auto-generated by scripts/gen-kernels-package.ts — DO NOT EDIT.
 // Self-contained runtime: the arrow functions below ARE the kernels.
 export const kernels = {
   "silu": {
@@ -29,11 +29,11 @@ export const kernels = {
   "gelu": {
     id: "gelu",
     precise:     {
-      js: "((x) => ((0.999993 * ((0.5 * x) * (1 + Math.tanh((0.797885 * (x + (0.044715 * (x) * (x) * (x)))))))) + -0.000058))",
-      eval: ((x) => ((0.999993 * ((0.5 * x) * (1 + Math.tanh((0.797885 * (x + (0.044715 * (x) * (x) * (x)))))))) + -0.000058)),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.999993f * ((0.5f * x) * (1f + tanhf((0.797885f * (x + (0.044715f * powf(x, 3.0f)))))))) + -0.000058f);\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.999993 * ((0.5 * x) * (1 + torch.tanh((0.797885 * (x + (0.044715 * (x ** 3)))))))) + -0.000058)",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgwBA2VudgR0YW5oAAEDAgEABwkBBXNwZWFyAAEKUAFOAERhU+dR8f/vP0QAAAAAAADgPyAAokQAAAAAAADwP0S4AZ8fRojpPyAARPcBSG3i5KY/IAAgACAAoqKioKIQAKCiokSQvknToGgOv6AL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return ((0.5 * x) * (1 + __erf((x * 0.7071067811865475)))); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return ((0.5 * x) * (1 + __erf((x * 0.7071067811865475)))); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.5f * x) * (1f + erff((x * 0.707107f))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.5 * x) * (1 + torch.erf((x * 0.707107))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQopAScARAAAAAAAAOA/IACiRAAAAAAAAPA/IABEzDt/Zp6g5j+iEACgogs=",
     },
     fast:     {
       js: "((x) => Math.max(0, x))",
@@ -43,11 +43,11 @@ export const kernels = {
       wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAAChABDgAgAEQAAAAAAAAAAKUL",
     },
     meta: {
-      metric: 3.55135869310835e-8,
+      metric: 1.5455587803707568e-34,
       level: 5,
-      formulaCost: 11,
+      formulaCost: 5,
       exactCost: 46,
-      speedupVsExact: 4.181818181818182,
+      speedupVsExact: 9.2,
       vsIterative: null,
     },
   },
@@ -919,25 +919,25 @@ export const kernels = {
   "erf_prob": {
     id: "erf_prob",
     precise:     {
-      js: "((x) => (1.106774 * Math.min(1e4, Math.max(-1e4, ((x + (0.034298 * (x) * (x) * (x)))) / (((0.995 + (0.378089 * (x) * (x)))) < 1e-4 && ((0.995 + (0.378089 * (x) * (x)))) > -1e-4 ? (((0.995 + (0.378089 * (x) * (x)))) >= 0 ? 1e-4 : -1e-4) : ((0.995 + (0.378089 * (x) * (x)))))))))",
-      eval: ((x) => (1.106774 * Math.min(1e4, Math.max(-1e4, ((x + (0.034298 * (x) * (x) * (x)))) / (((0.995 + (0.378089 * (x) * (x)))) < 1e-4 && ((0.995 + (0.378089 * (x) * (x)))) > -1e-4 ? (((0.995 + (0.378089 * (x) * (x)))) >= 0 ? 1e-4 : -1e-4) : ((0.995 + (0.378089 * (x) * (x))))))))),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (1.106774f * fminf(fmaxf(((x + (0.034298f * powf(x, 3.0f))) / copysignf(fmaxf(fabsf((0.995f + (0.378089f * powf(x, 2.0f)))), 1.0e-4f), (0.995f + (0.378089f * powf(x, 2.0f))))), -1.0e4f), 1.0e4f));\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (1.106774 * ((x + (0.034298 * (x ** 3))) / (abs((0.995 + (0.378089 * (x * x)))) < 1e-4 ? 1e-4 : (0.995 + (0.378089 * (x * x))))))",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACnYBdABEeQJhp1i18T8gAEQp7KLogY+hPyAAIAAgAKKioqBE16NwPQrX7z9Ezox+NJwy2D8gACAAoqKgmUQtQxzr4jYaP6VE16NwPQrX7z9Ezox+NJwy2D8gACAAoqKgpqNEAAAAAACIw8ClRAAAAAAAiMNApKIL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return erff(x);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.erf(x)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQoIAQYAIAAQAAs=",
     },
     fast:     {
-      js: "((x) => Math.atan(x))",
-      eval: ((x) => Math.atan(x)),
-      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return atanf(x);\n}",
-      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.atan(x)",
-      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgwBA2VudgRhdGFuAAEDAgEABwkBBXNwZWFyAAEKCAEGACAAEAAL",
+      js: "((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); })",
+      eval: ((x) => { const __erf = (x) => { const s = x < 0 ? -1 : 1; const ax = Math.abs(x); const t = 1 / (1 + 0.3275911 * ax); return s * (1 - ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * Math.exp(-ax * ax)); }; return __erf(x); }),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return erff(x);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.erf(x)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNlcmYAAQMCAQAHCQEFc3BlYXIAAQoIAQYAIAAQAAs=",
     },
     meta: {
-      metric: 0.00004650870031371604,
-      level: 3,
-      formulaCost: 12,
+      metric: 0,
+      level: 5,
+      formulaCost: 1,
       exactCost: 26,
-      speedupVsExact: 2.1666666666666665,
+      speedupVsExact: 26,
       vsIterative: null,
     },
   },
@@ -2002,8 +2002,158 @@ export const kernels = {
       vsIterative: null,
     },
   },
+  "rope_freq": {
+    id: "rope_freq",
+    precise:     {
+      js: "((x) => ((0.9999991529919943 * ((0.999947 * Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))) + -0.000038)) + 0.00001879512303862185))",
+      eval: ((x) => ((0.9999991529919943 * ((0.999947 * Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))) + -0.000038)) + 0.00001879512303862185)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.999999f * ((0.999947f * cosf((1023.994f * expf(fmaxf(-50.0f, fminf(50.0f, (-0.28782f * x))))))) + -0.000038f)) + 0.000019f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.999999 * ((0.999947 * torch.cos((1023.994 * torch.exp(torch.clamp((-0.28782 * x), -50.0, 50.0))))) + -0.000038)) + 0.000019)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AhUCA2VudgNleHAAAQNlbnYDY29zAAEDAgEABwkBBXNwZWFyAAIKWgFYAEQqG0Q5/v/vP0RLd9fZkP/vP0TLoUW28/+PQERBn8iTpGvSvyAAokQAAAAAAABJwKVEAAAAAAAASUCkEACiEAGiRBgK2A5G7AO/oKJEuB45BUe18z6gCw==",
+    },
+    fast:     {
+      js: "((x) => Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x)))))))",
+      eval: ((x) => Math.cos((1023.994 * Math.exp(Math.max(-50, Math.min(50, (-0.28782 * x))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return cosf((1023.994f * expf(fmaxf(-50.0f, fminf(50.0f, (-0.28782f * x))))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.cos((1023.994 * torch.exp(torch.clamp((-0.28782 * x), -50.0, 50.0))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AhUCA2VudgNleHAAAQNlbnYDY29zAAEDAgEABwkBBXNwZWFyAAIKMgEwAETLoUW28/+PQERBn8iTpGvSvyAAokQAAAAAAABJwKVEAAAAAAAASUCkEACiEAEL",
+    },
+    meta: {
+      metric: 8.450097215964813e-7,
+      level: 5,
+      formulaCost: 8,
+      exactCost: 43,
+      speedupVsExact: 5.375,
+      vsIterative: null,
+    },
+  },
+  "bilateral_weight": {
+    id: "bilateral_weight",
+    precise:     {
+      js: "((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2)))))))",
+      eval: ((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return fminf(fmaxf((expf(fmaxf(-50.0f, fminf(50.0f, (-(powf(x, 2.0f) * 50f))))) / copysignf(fmaxf(fabsf((1f + (powf(x, 2.0f) * 2f))), 1.0e-4f), (1f + (powf(x, 2.0f) * 2f)))), -1.0e4f), 1.0e4f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (torch.exp(torch.clamp((-((x * x) * 50)), -50.0, 50.0)) / (abs((1 + ((x * x) * 2))) < 1e-4 ? 1e-4 : (1 + ((x * x) * 2))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNleHAAAQMCAQAHCQEFc3BlYXIAAQp9AXsAIAAgAKJEAAAAAAAASUCimkQAAAAAAABJwKVEAAAAAAAASUCkEABEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgmUQtQxzr4jYaP6VEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgpqNEAAAAAACIw8ClRAAAAAAAiMNApAs=",
+    },
+    fast:     {
+      js: "((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2)))))))",
+      eval: ((x) => Math.min(1e4, Math.max(-1e4, (Math.exp(Math.max(-50, Math.min(50, (-((x) * (x) * 50)))))) / (((1 + ((x) * (x) * 2))) < 1e-4 && ((1 + ((x) * (x) * 2))) > -1e-4 ? (((1 + ((x) * (x) * 2))) >= 0 ? 1e-4 : -1e-4) : ((1 + ((x) * (x) * 2))))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return fminf(fmaxf((expf(fmaxf(-50.0f, fminf(50.0f, (-(powf(x, 2.0f) * 50f))))) / copysignf(fmaxf(fabsf((1f + (powf(x, 2.0f) * 2f))), 1.0e-4f), (1f + (powf(x, 2.0f) * 2f)))), -1.0e4f), 1.0e4f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (torch.exp(torch.clamp((-((x * x) * 50)), -50.0, 50.0)) / (abs((1 + ((x * x) * 2))) < 1e-4 ? 1e-4 : (1 + ((x * x) * 2))))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNleHAAAQMCAQAHCQEFc3BlYXIAAQp9AXsAIAAgAKJEAAAAAAAASUCimkQAAAAAAABJwKVEAAAAAAAASUCkEABEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgmUQtQxzr4jYaP6VEAAAAAAAA8D8gACAAokQAAAAAAAAAQKKgpqNEAAAAAACIw8ClRAAAAAAAiMNApAs=",
+    },
+    meta: {
+      metric: 9.336399957342636e-35,
+      level: 5,
+      formulaCost: 11,
+      exactCost: 45,
+      speedupVsExact: 4.090909090909091,
+      vsIterative: null,
+    },
+  },
+  "bs_d1_sigma": {
+    id: "bs_d1_sigma",
+    precise:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((Math.abs(x)) < 1e-4 && (Math.abs(x)) > -1e-4 ? ((Math.abs(x)) >= 0 ? 1e-4 : -1e-4) : (Math.abs(x))))) + (0.25 * Math.abs(x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((Math.abs(x)) < 1e-4 && (Math.abs(x)) > -1e-4 ? ((Math.abs(x)) >= 0 ? 1e-4 : -1e-4) : (Math.abs(x))))) + (0.25 * Math.abs(x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(fabsf(x)), 1.0e-4f), fabsf(x))), -1.0e4f), 1.0e4f) + (0.25f * fabsf(x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(torch.abs(x)) < 1e-4 ? 1e-4 : torch.abs(x))) + (0.25 * torch.abs(x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkIBQABEop9GrXKZyz8gAJmZRC1DHOviNho/pSAAmaajRAAAAAAAiMPApUQAAAAAAIjDQKREAAAAAAAA0D8gAJmioAs=",
+    },
+    fast:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (3.635301) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (4.215319 * x)))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (3.635301) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (4.215319 * x))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((3.635301f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (4.215319f * x));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((3.635301 / (abs(x) < 1e-4 ? 1e-4 : x)) + (4.215319 * x))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACj8BPQBExO3QsBgVDUAgAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRGfUfJV83BBAIACioAs=",
+    },
+    meta: {
+      metric: 1.6217711315014304e-18,
+      level: 5,
+      formulaCost: 8,
+      exactCost: 9,
+      speedupVsExact: 1.125,
+      vsIterative: null,
+    },
+  },
+  "mel_scale": {
+    id: "mel_scale",
+    precise:     {
+      js: "((x) => (1126.994181 * Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286))))))",
+      eval: ((x) => (1126.994181 * Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286)))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (1126.994181f * logf(fmaxf(1.0e-30f, (1f + (x * 0.001429f)))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return (1126.994181 * torch.log(torch.clamp((1 + (x * 0.001429)), min=1e-30)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQowAS4ARDiFlQr6m5FARAAAAAAAAPA/IABEsak05NxnVz+ioESgwuv+S0i0OaUQAKIL",
+    },
+    fast:     {
+      js: "((x) => Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286)))))",
+      eval: ((x) => Math.log(Math.max(1e-30, (1 + (x * 0.0014285714285714286))))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return logf(fmaxf(1.0e-30f, (1f + (x * 0.001429f))));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return torch.log(torch.clamp((1 + (x * 0.001429)), min=1e-30))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQomASQARAAAAAAAAPA/IABEsak05NxnVz+ioESgwuv+S0i0OaUQAAs=",
+    },
+    meta: {
+      metric: 1.4594066099947563e-12,
+      level: 5,
+      formulaCost: 4,
+      exactCost: 23,
+      speedupVsExact: 5.75,
+      vsIterative: null,
+    },
+  },
+  "bs_d2_sigma": {
+    id: "bs_d2_sigma",
+    precise:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (-(0.25f * x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(x) < 1e-4 ? 1e-4 : x)) + (-(0.25 * x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkABPgBEop9GrXKZyz8gAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRAAAAAAAANA/IACimqAL",
+    },
+    fast:     {
+      js: "((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x))))",
+      eval: ((x) => (Math.min(1e4, Math.max(-1e4, (0.21562036) / ((x) < 1e-4 && (x) > -1e-4 ? ((x) >= 0 ? 1e-4 : -1e-4) : (x)))) + (-(0.25 * x)))),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return (fminf(fmaxf((0.21562f / copysignf(fmaxf(fabsf(x), 1.0e-4f), x)), -1.0e4f), 1.0e4f) + (-(0.25f * x)));\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.21562 / (abs(x) < 1e-4 ? 1e-4 : x)) + (-(0.25 * x)))",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AwIBAAcJAQVzcGVhcgAACkABPgBEop9GrXKZyz8gAJlELUMc6+I2Gj+lIACmo0QAAAAAAIjDwKVEAAAAAACIw0CkRAAAAAAAANA/IACimqAL",
+    },
+    meta: {
+      metric: 1.6217711130493227e-18,
+      level: 5,
+      formulaCost: 7,
+      exactCost: 11,
+      speedupVsExact: 1.5714285714285714,
+      vsIterative: null,
+    },
+  },
+  "a_weighting": {
+    id: "a_weighting",
+    precise:     {
+      js: "((x) => ((0.9803615530479888 * ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)) + 0.039276893904016105))",
+      eval: ((x) => ((0.9803615530479888 * ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)) + 0.039276893904016105)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((0.980362f * ((8.859884f * logf(fmaxf(1.0e-30f, fminf(fmaxf((((148691236f * powf(x, 2.0f)) * powf(x, 2.0f)) / copysignf(fmaxf(fabsf(((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f)))), 1.0e-4f), ((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f))))), -1.0e4f), 1.0e4f)))) + 2f)) + 0.039277f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((0.980362 * ((8.859884 * torch.log(torch.clamp((((148691236 * (x * x)) * (x * x)) / (abs((((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236)))) < 1e-4 ? 1e-4 : (((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236))))), min=1e-30))) + 2)) + 0.039277)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQrxAQHuAQBEphkTMR9f7z9EE7U0t0K4IUBEAAAASLK5oUEgACAAoqIgACAAoqIgACAAokT2KFyPwoV6QKAgACAAokTsUbgepafGQKAgACAAokQfhevR4J0gQaCimZ8gACAAokQAAABIsrmhQaCioplELUMc6+I2Gj+lIAAgAKJE9ihcj8KFekCgIAAgAKJE7FG4HqWnxkCgIAAgAKJEH4Xr0eCdIEGgopmfIAAgAKJEAAAASLK5oUGgoqKmo0QAAAAAAIjDwKVEAAAAAACIw0CkRKDC6/5LSLQ5pRAAokQAAAAAAAAAQKCiRK7HnN0ZHKQ/oAs=",
+    },
+    fast:     {
+      js: "((x) => ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2))",
+      eval: ((x) => ((8.859884 * Math.log(Math.max(1e-30, Math.min(1e4, Math.max(-1e4, (((148691236 * (x) * (x)) * (x) * (x))) / (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) < 1e-4 && ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) > -1e-4 ? (((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))) >= 0 ? 1e-4 : -1e-4) : ((((x) * (x) + 424.36) * (Math.sqrt(Math.abs((((x) * (x) + 11599.29) * ((x) * (x) + 544496.41)))) * ((x) * (x) + 148691236)))))))))) + 2)),
+      c: "// Evolved by SPEAR — algebraic only, FP16 safe\n__device__ inline float spear_fn(const float x) {\n    return ((8.859884f * logf(fmaxf(1.0e-30f, fminf(fmaxf((((148691236f * powf(x, 2.0f)) * powf(x, 2.0f)) / copysignf(fmaxf(fabsf(((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f)))), 1.0e-4f), ((powf(x, 2.0f) + 424.36f) * (sqrtf(fabsf(((powf(x, 2.0f) + 11599.29f) * (powf(x, 2.0f) + 544496.41f)))) * (powf(x, 2.0f) + 148691236f))))), -1.0e4f), 1.0e4f)))) + 2f);\n}",
+      py: "import torch\n\ndef spear_fn(x):\n    # Evolved by SPEAR — zero transcendental ops (exp/erf/tanh free)\n    return ((8.859884 * torch.log(torch.clamp((((148691236 * (x * x)) * (x * x)) / (abs((((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236)))) < 1e-4 ? 1e-4 : (((x * x) + 424.36) * (torch.sqrt(torch.abs((((x * x) + 11599.29) * ((x * x) + 544496.41)))) * ((x * x) + 148691236))))), min=1e-30))) + 2)",
+      wasmBase64: "AGFzbQEAAAABCwJgAXwBfGABfAF8AgsBA2VudgNsb2cAAQMCAQAHCQEFc3BlYXIAAQrdAQHaAQBEE7U0t0K4IUBEAAAASLK5oUEgACAAoqIgACAAoqIgACAAokT2KFyPwoV6QKAgACAAokTsUbgepafGQKAgACAAokQfhevR4J0gQaCimZ8gACAAokQAAABIsrmhQaCioplELUMc6+I2Gj+lIAAgAKJE9ihcj8KFekCgIAAgAKJE7FG4HqWnxkCgIAAgAKJEH4Xr0eCdIEGgopmfIAAgAKJEAAAASLK5oUGgoqKmo0QAAAAAAIjDwKVEAAAAAACIw0CkRKDC6/5LSLQ5pRAAokQAAAAAAAAAQKAL",
+    },
+    meta: {
+      metric: 2.7800484820648814e-29,
+      level: 5,
+      formulaCost: 26,
+      exactCost: 55,
+      speedupVsExact: 2.1153846153846154,
+      vsIterative: null,
+    },
+  },
 };
 
-export const kernelIds = ["silu","gelu","sigmoid","kv_cache","free_fall","kepler","gaussian_cdf","damped_pendulum","rl_distillation","lambert_w","rc_circuit","gaussian_kernel","diffusion_beta","bilinear_interp","european_call","temporal_grad","lorentz","hill","kerr","lennard_jones","damped_oscillation","logistic_growth","softplus","kdv_soliton","kerr_spin","pendulum_hybrid","eigen3_sym","ik_reach","idm_following","gemv4","rope_rot","gauss_shader","ema_smooth","smoothstep","srgb_gamma","tanh_sat","atan_unit","srgb_decode","erf_prob","huber_loss","cosh_curve","bessel_j0","bessel_j1","blackbody_r","aces_fit","logsumexp2","probit_quantile","pmt_finance","blackbody_g","blackbody_b","bessel_j2","logit_ml","kelly_criterion","rsi_momentum","implied_vol","michaelis_menten","temperature_softmax","doppler_effect","stefan_boltzmann","mm1_queue_wait","grover_amplitude","concurrence_pure","chsh_correlation","legendre_p2","laguerre_l2","asin_hard","qfi_dephasing","amp_damp_fid","loschmidt_rate","bessel_i0e","elliptic_k","kepler_solver","fast_exp_alu","fresnel_schlick","rayleigh_phase","smootherstep","fog_exp2","back_ease_out","light_falloff_punctual","bias_slope","uncharted2_tonemap","mish"];
+export const kernelIds = ["silu","gelu","sigmoid","kv_cache","free_fall","kepler","gaussian_cdf","damped_pendulum","rl_distillation","lambert_w","rc_circuit","gaussian_kernel","diffusion_beta","bilinear_interp","european_call","temporal_grad","lorentz","hill","kerr","lennard_jones","damped_oscillation","logistic_growth","softplus","kdv_soliton","kerr_spin","pendulum_hybrid","eigen3_sym","ik_reach","idm_following","gemv4","rope_rot","gauss_shader","ema_smooth","smoothstep","srgb_gamma","tanh_sat","atan_unit","srgb_decode","erf_prob","huber_loss","cosh_curve","bessel_j0","bessel_j1","blackbody_r","aces_fit","logsumexp2","probit_quantile","pmt_finance","blackbody_g","blackbody_b","bessel_j2","logit_ml","kelly_criterion","rsi_momentum","implied_vol","michaelis_menten","temperature_softmax","doppler_effect","stefan_boltzmann","mm1_queue_wait","grover_amplitude","concurrence_pure","chsh_correlation","legendre_p2","laguerre_l2","asin_hard","qfi_dephasing","amp_damp_fid","loschmidt_rate","bessel_i0e","elliptic_k","kepler_solver","fast_exp_alu","fresnel_schlick","rayleigh_phase","smootherstep","fog_exp2","back_ease_out","light_falloff_punctual","bias_slope","uncharted2_tonemap","mish","rope_freq","bilateral_weight","bs_d1_sigma","mel_scale","bs_d2_sigma","a_weighting"];
 
 export default kernels;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "spear-kernels",
-  "version": "1.7.0",
+  "version": "1.9.0",
   "description": "Closed-form kernels discovered by the SPEAR symbolic regression engine, shipped as JS, CUDA C, PyTorch and WebAssembly.",
   "license": "MIT",
   "type": "module",