spear-kernels

Every closed-form kernel discovered by the SPEAR symbolic regression engine, published as a zero-dependency, fully self-contained package. Each kernel ships four interchangeable backends plus metadata:

| Field | What it is | |---|---| | precise.js / precise.eval | standalone JS arrow-function (source string / compiled) | | precise.c | CUDA C __device__ source | | precise.py | PyTorch source | | precise.wasmBase64 | compiled f64 WebAssembly module | | fast | algebraic-only variant when one was discovered | | meta | metric, level, formulaCost, exactCost, speedupVsExact, vsIterative |

Usage

import { kernels, kernelIds } from "spear-kernels";

kernelIds;                       // list of all kernel ids

kernels.silu.precise.eval(2);    // ~1.762 (SiLU approximant)

if (kernels.sigmoid.fast) {
  kernels.sigmoid.fast.eval(3);  // algebraic fast path
}

kernels.gaussian_cdf.meta.speedupVsExact; // metadata access

WebAssembly backend

const bytes = Uint8Array.from(atob(kernels.silu.precise.wasmBase64), (c) => c.charCodeAt(0));
const env = {
  exp: Math.exp, sin: Math.sin, cos: Math.cos, atan: Math.atan,
  log: (v) => Math.log(v > 1e-30 ? v : 1e-30),
};
const { instance } = await WebAssembly.instantiate(bytes, { env });
instance.exports.spear(2); // same protected-division semantics as the JS backend

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 /.

Support

If SPEAR kernels save you compute time, consider tipping the lab:

₿ bc1q54n9x2894rr43f7nkaqywtegn8aufxwlnnkh5r

npm fund also surfaces this address (declared in the funding field). Every satoshi goes back into GPU time for the evolution farm.