@seanhogg/builderforce-memory-engine 2026.9.34 → 2026.10.2

package/dist/ml/classical/math.js

@@ -0,0 +1,90 @@
+/**
+ * ml/classical/math.ts — the small vector helpers every classical algorithm shares.
+ * Pure, allocation-light, deterministic.
+ */
+import { SeededRng } from "../../utils/rng.js";
+/** The engine's ONE sigmoid / softmax (`training/cpu_ops.ts`). */
+export { sigmoid, softmax } from "../../training/cpu_ops.js";
+export function dot(a, b) {
+    const n = Math.min(a.length, b.length);
+    let s = 0;
+    for (let i = 0; i < n; i++)
+        s += a[i] * b[i];
+    return s;
+}
+export function norm(a) {
+    return Math.sqrt(dot(a, a));
+}
+/** Cosine similarity in [-1, 1]; 0 when either vector is zero. */
+export function cosine(a, b) {
+    const na = norm(a);
+    const nb = norm(b);
+    return na > 0 && nb > 0 ? dot(a, b) / (na * nb) : 0;
+}
+export function squaredDistance(a, b) {
+    const n = Math.min(a.length, b.length);
+    let s = 0;
+    for (let i = 0; i < n; i++) {
+        const d = a[i] - b[i];
+        s += d * d;
+    }
+    return s;
+}
+/** Median of a numeric list (copy-sorted). NaN for an empty list. */
+export function median(values) {
+    if (values.length === 0)
+        return Number.NaN;
+    const s = [...values].sort((a, b) => a - b);
+    const m = s.length >> 1;
+    return s.length % 2 ? s[m] : (s[m - 1] + s[m]) / 2;
+}
+/** Index of the largest value (first on ties). */
+export function argmax(values) {
+    let best = 0;
+    for (let i = 1; i < values.length; i++)
+        if (values[i] > values[best])
+            best = i;
+    return best;
+}
+/** Number of distinct classes implied by integer labels 0..k-1. */
+export function classCount(labels) {
+    let k = 0;
+    for (const y of labels) {
+        if (!Number.isInteger(y) || y < 0)
+            throw new Error(`class label must be a non-negative integer, got ${y}`);
+        k = Math.max(k, y + 1);
+    }
+    return Math.max(k, 2);
+}
+/** Feature dimension of a non-empty sample matrix; throws on ragged rows. */
+export function featureDim(X) {
+    if (X.length === 0)
+        throw new Error("no samples");
+    const d = X[0].length;
+    for (const x of X)
+        if (x.length !== d)
+            throw new Error(`ragged features: expected ${d}, got ${x.length}`);
+    return d;
+}
+/** Confidence of a probability vector: its top probability. */
+export function topProbability(p) {
+    return p.length ? Math.max(...p) : 0;
+}
+/** A deterministic [0,1) source for algorithms that sample — the engine's ONE seeded RNG. */
+export function seededUniform(seed) {
+    const rng = new SeededRng(seed);
+    return () => rng.next();
+}
+/**
+ * Average path length of an unsuccessful BST search over n points — the
+ * isolation-forest normaliser c(n).
+ */
+export function averagePathLength(n) {
+    if (n <= 1)
+        return 0;
+    if (n === 2)
+        return 1;
+    const harmonic = Math.log(n - 1) + 0.5772156649;
+    return 2 * harmonic - (2 * (n - 1)) / n;
+}
+//# sourceMappingURL=math.js.map

package/dist/ml/classical/math.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"math.js","sourceRoot":"","sources":["../../../src/ml/classical/math.ts"],"names":[],"mappings":"AAAA;;;GAGG;AACH,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAC;AAE/C,kEAAkE;AAClE,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,2BAA2B,CAAC;AAI7D,MAAM,UAAU,GAAG,CAAC,CAAM,EAAE,CAAM;IAChC,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC;IACvC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,CAAC,IAAI,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;IAC/C,OAAO,CAAC,CAAC;AACX,CAAC;AAED,MAAM,UAAU,IAAI,CAAC,CAAM;IACzB,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;AAC9B,CAAC;AAED,kEAAkE;AAClE,MAAM,UAAU,MAAM,CAAC,CAAM,EAAE,CAAM;IACnC,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;IACnB,MAAM,EAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;IACnB,OAAO,EAAE,GAAG,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACtD,CAAC;AAED,MAAM,UAAU,eAAe,CAAC,CAAM,EAAE,CAAM;IAC5C,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC;IACvC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3B,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;QACxB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACb,CAAC;IACD,OAAO,CAAC,CAAC;AACX,CAAC;AAED,qEAAqE;AACrE,MAAM,UAAU,MAAM,CAAC,MAAyB;IAC9C,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC,GAAG,CAAC;IAC3C,MAAM,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IAC5C,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC;IACxB,OAAO,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,GAAG,CAAC,CAAC;AACxD,CAAC;AAED,kDAAkD;AAClD,MAAM,UAAU,MAAM,CAAC,MAAyB;IAC9C,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,IAAI,MAAM,CAAC,CAAC,CAAE,GAAG,MAAM,CAAC,IAAI,CAAE;YAAE,IAAI,GAAG,CAAC,CAAC;IACjF,OAAO,IAAI,CAAC;AACd,CAAC;AAED,mEAAmE;AACnE,MAAM,UAAU,UAAU,CAAC,MAAyB;IAClD,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;QACvB,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,mDAAmD,CAAC,EAAE,CAAC,CAAC;QAC3G,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC;IACzB,CAAC;IACD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;AACxB,CAAC;AAED,6EAA6E;AAC7E,MAAM,UAAU,UAAU,CAAC,CAAiB;IAC1C,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,YAAY,CAAC,CAAC;IAClD,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,MAAM,CAAC;IACvB,KAAK,MAAM,CAAC,IAAI,CAAC;QAAE,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,6BAA6B,CAAC,SAAS,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;IAC1G,OAAO,CAAC,CAAC;AACX,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,cAAc,CAAC,CAAoB;IACjD,OAAO,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACvC,CAAC;AAED,6FAA6F;AAC7F,MAAM,UAAU,aAAa,CAAC,IAAY;IACxC,MAAM,GAAG,GAAG,IAAI,SAAS,CAAC,IAAI,CAAC,CAAC;IAChC,OAAO,GAAG,EAAE,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC;AAC1B,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,CAAS;IACzC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,CAAC,CAAC;IACrB,IAAI,CAAC,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IACtB,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,YAAY,CAAC;IAChD,OAAO,CAAC,GAAG,QAAQ,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;AAC1C,CAAC"}

package/dist/ml/classical/naive_bayes.d.ts

@@ -0,0 +1,28 @@
+/**
+ * ml/classical/naive_bayes.ts — Gaussian naive Bayes.
+ *
+ * Per class: a weighted prior, and an independent Gaussian per feature. Variance
+ * is smoothed by a fraction of the largest feature variance (as scikit-learn's
+ * `var_smoothing`), so a constant feature never divides by zero.
+ *
+ * A class with no training weight has prior −∞ and probability exactly 0: its
+ * Gaussian is a placeholder (mean 0, variance ≈ the smoothing floor), which would
+ * otherwise out-score every real class on an input near zero.
+ */
+import type { MlPrediction, MlPredictor, MlSample } from "../types.js";
+export interface NaiveBayesOptions {
+    classes?: number;
+    /** Fraction of the largest feature variance added to every variance. Default 1e-9. */
+    varSmoothing?: number;
+}
+export declare class NaiveBayesModel implements MlPredictor {
+    readonly logPrior: Float64Array;
+    readonly mean: Float64Array[];
+    readonly variance: Float64Array[];
+    constructor(logPrior: Float64Array, mean: Float64Array[], variance: Float64Array[]);
+    /** Per-class joint log-likelihood log P(c) + Σ log N(x_i | μ, σ²); −∞ for an unseen class. */
+    jointLogLikelihood(x: ArrayLike<number>): number[];
+    predict(x: Float32Array): MlPrediction;
+}
+export declare function fitNaiveBayes(samples: readonly MlSample[], opts?: NaiveBayesOptions): NaiveBayesModel;
+//# sourceMappingURL=naive_bayes.d.ts.map

package/dist/ml/classical/naive_bayes.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"naive_bayes.d.ts","sourceRoot":"","sources":["../../../src/ml/classical/naive_bayes.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAGvE,MAAM,WAAW,iBAAiB;IAChC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,sFAAsF;IACtF,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,qBAAa,eAAgB,YAAW,WAAW;IAE/C,QAAQ,CAAC,QAAQ,EAAE,YAAY;IAC/B,QAAQ,CAAC,IAAI,EAAE,YAAY,EAAE;IAC7B,QAAQ,CAAC,QAAQ,EAAE,YAAY,EAAE;gBAFxB,QAAQ,EAAE,YAAY,EACtB,IAAI,EAAE,YAAY,EAAE,EACpB,QAAQ,EAAE,YAAY,EAAE;IAGnC,8FAA8F;IAC9F,kBAAkB,CAAC,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,GAAG,MAAM,EAAE;IAalD,OAAO,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY;CAIvC;AAED,wBAAgB,aAAa,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE,EAAE,IAAI,GAAE,iBAAsB,GAAG,eAAe,CAoCzG"}

package/dist/ml/classical/naive_bayes.js

@@ -0,0 +1,72 @@
+import { classCount, featureDim, softmax, topProbability } from "./math.js";
+export class NaiveBayesModel {
+    logPrior;
+    mean;
+    variance;
+    constructor(logPrior, mean, variance) {
+        this.logPrior = logPrior;
+        this.mean = mean;
+        this.variance = variance;
+    }
+    /** Per-class joint log-likelihood log P(c) + Σ log N(x_i | μ, σ²); −∞ for an unseen class. */
+    jointLogLikelihood(x) {
+        return this.mean.map((mu, c) => {
+            const v = this.variance[c];
+            let ll = this.logPrior[c];
+            if (ll === -Infinity)
+                return ll;
+            for (let i = 0; i < mu.length; i++) {
+                const d = (x[i] ?? 0) - mu[i];
+                ll += -0.5 * Math.log(2 * Math.PI * v[i]) - (d * d) / (2 * v[i]);
+            }
+            return ll;
+        });
+    }
+    predict(x) {
+        const scores = Array.from(softmax(this.jointLogLikelihood(x)));
+        return { scores, confidence: topProbability(scores) };
+    }
+}
+export function fitNaiveBayes(samples, opts = {}) {
+    const d = featureDim(samples.map((s) => s.x));
+    const k = opts.classes ?? classCount(samples.map((s) => s.y));
+    const wSum = new Float64Array(k);
+    const mean = Array.from({ length: k }, () => new Float64Array(d));
+    const variance = Array.from({ length: k }, () => new Float64Array(d));
+    for (const s of samples) {
+        const w = s.weight ?? 1;
+        wSum[s.y] = wSum[s.y] + w;
+        for (let i = 0; i < d; i++)
+            mean[s.y][i] = mean[s.y][i] + w * s.x[i];
+    }
+    for (let c = 0; c < k; c++)
+        if (wSum[c] > 0)
+            for (let i = 0; i < d; i++)
+                mean[c][i] = mean[c][i] / wSum[c];
+    for (const s of samples) {
+        const w = s.weight ?? 1;
+        for (let i = 0; i < d; i++) {
+            const diff = s.x[i] - mean[s.y][i];
+            variance[s.y][i] = variance[s.y][i] + w * diff * diff;
+        }
+    }
+    let maxVar = 0;
+    for (let c = 0; c < k; c++) {
+        for (let i = 0; i < d; i++) {
+            variance[c][i] = wSum[c] > 0 ? variance[c][i] / wSum[c] : 0;
+            maxVar = Math.max(maxVar, variance[c][i]);
+        }
+    }
+    const eps = (opts.varSmoothing ?? 1e-9) * Math.max(maxVar, 1e-12) + 1e-12;
+    for (let c = 0; c < k; c++)
+        for (let i = 0; i < d; i++)
+            variance[c][i] = variance[c][i] + eps;
+    const total = wSum.reduce((a, b) => a + b, 0);
+    if (!(total > 0))
+        throw new Error("fitNaiveBayes: the samples carry no weight");
+    // An unseen class is masked out (−∞, so softmax gives it exactly 0). At least one
+    // class has weight, so softmax always has a finite maximum.
+    const logPrior = Float64Array.from(wSum, (w) => (w > 0 ? Math.log(w / total) : -Infinity));
+    return new NaiveBayesModel(logPrior, mean, variance);
+}
+//# sourceMappingURL=naive_bayes.js.map

package/dist/ml/classical/naive_bayes.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"naive_bayes.js","sourceRoot":"","sources":["../../../src/ml/classical/naive_bayes.ts"],"names":[],"mappings":"AAYA,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,OAAO,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAQ5E,MAAM,OAAO,eAAe;IAEf;IACA;IACA;IAHX,YACW,QAAsB,EACtB,IAAoB,EACpB,QAAwB;QAFxB,aAAQ,GAAR,QAAQ,CAAc;QACtB,SAAI,GAAJ,IAAI,CAAgB;QACpB,aAAQ,GAAR,QAAQ,CAAgB;IAChC,CAAC;IAEJ,8FAA8F;IAC9F,kBAAkB,CAAC,CAAoB;QACrC,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,EAAE;YAC7B,MAAM,CAAC,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC;YAC5B,IAAI,EAAE,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC;YAC3B,IAAI,EAAE,KAAK,CAAC,QAAQ;gBAAE,OAAO,EAAE,CAAC;YAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,EAAE,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACnC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,CAAC;gBAC/B,EAAE,IAAI,CAAC,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;YACrE,CAAC;YACD,OAAO,EAAE,CAAC;QACZ,CAAC,CAAC,CAAC;IACL,CAAC;IAED,OAAO,CAAC,CAAe;QACrB,MAAM,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC/D,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC;IACxD,CAAC;CACF;AAED,MAAM,UAAU,aAAa,CAAC,OAA4B,EAAE,OAA0B,EAAE;IACtF,MAAM,CAAC,GAAG,UAAU,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC9C,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,IAAI,UAAU,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC9D,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IACjC,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,EAAE,GAAG,EAAE,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC;IAClE,MAAM,QAAQ,GAAG,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,EAAE,GAAG,EAAE,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC;IAEtE,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC;QACxB,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC;QAC3B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YAAE,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC;IAC3E,CAAC;IACD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,IAAI,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC;YAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;gBAAE,IAAI,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;IAChH,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC;QACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,CAAC;YACtC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,GAAG,CAAC,GAAG,IAAI,GAAG,IAAI,CAAC;QAC3D,CAAC;IACH,CAAC;IACD,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC,CAAC;YACjE,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,CAAC,CAAC;QAC9C,CAAC;IACH,CAAC;IACD,MAAM,GAAG,GAAG,CAAC,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,KAAK,CAAC,GAAG,KAAK,CAAC;IAC1E,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;YAAE,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAE,GAAG,GAAG,CAAC;IAEjG,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC;IAC9C,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,4CAA4C,CAAC,CAAC;IAChF,kFAAkF;IAClF,4DAA4D;IAC5D,MAAM,QAAQ,GAAG,YAAY,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC;IAC3F,OAAO,IAAI,eAAe,CAAC,QAAQ,EAAE,IAAI,EAAE,QAAQ,CAAC,CAAC;AACvD,CAAC"}

package/dist/ml/classical/robust_z.d.ts

@@ -0,0 +1,38 @@
+/**
+ * ml/classical/robust_z.ts — robust (median/MAD) z-scores for anomaly detection.
+ *
+ * Per feature: z = |x − median| / (1.4826 · MAD). Median and MAD are not dragged
+ * by the very outliers being looked for, unlike mean and standard deviation. When
+ * MAD is 0 (more than half the history identical) the mean absolute deviation
+ * stands in, then two floors: a fraction of |median| (relative) and `minScale`
+ * (absolute, in the feature's own units). Without the absolute one a history that
+ * is all zeros — no tool failures, ever — gets a scale of ~1e-9, and the first
+ * small non-zero value scores astronomically.
+ *
+ * Prediction: `scores[0] = max_i z_i / threshold` — ≥ 1 is anomalous.
+ */
+import type { MlPrediction, MlPredictor, MlSample } from "../types.js";
+export interface RobustZOptions {
+    /** z at which a feature counts as anomalous. Default 3.5 (Iglewicz & Hoaglin). */
+    threshold?: number;
+    /** Smallest scale per feature, as a fraction of |median| (plus 1e-9). Default 0.01. */
+    minScaleFraction?: number;
+    /**
+     * Smallest scale per feature in absolute feature units. Default 0 — the algorithm
+     * cannot know the units; a task on a fixed scale sets it (`run.anomaly` does).
+     */
+    minScale?: number;
+}
+export declare class RobustZModel implements MlPredictor {
+    readonly center: Float64Array;
+    readonly scale: Float64Array;
+    readonly threshold: number;
+    constructor(center: Float64Array, scale: Float64Array, threshold: number);
+    /** Per-feature robust z-scores (absolute). */
+    zScores(x: ArrayLike<number>): number[];
+    predict(x: Float32Array): MlPrediction;
+}
+/** Confidence of an anomaly score normalised so 1 is the threshold: distance from it, capped. */
+export declare function anomalyConfidence(score: number): number;
+export declare function fitRobustZ(samples: readonly MlSample[], opts?: RobustZOptions): RobustZModel;
+//# sourceMappingURL=robust_z.d.ts.map

package/dist/ml/classical/robust_z.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"robust_z.d.ts","sourceRoot":"","sources":["../../../src/ml/classical/robust_z.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAQvE,MAAM,WAAW,cAAc;IAC7B,kFAAkF;IAClF,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,uFAAuF;IACvF,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B;;;OAGG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,qBAAa,YAAa,YAAW,WAAW;IAE5C,QAAQ,CAAC,MAAM,EAAE,YAAY;IAC7B,QAAQ,CAAC,KAAK,EAAE,YAAY;IAC5B,QAAQ,CAAC,SAAS,EAAE,MAAM;gBAFjB,MAAM,EAAE,YAAY,EACpB,KAAK,EAAE,YAAY,EACnB,SAAS,EAAE,MAAM;IAG5B,8CAA8C;IAC9C,OAAO,CAAC,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,GAAG,MAAM,EAAE;IAIvC,OAAO,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY;CAIvC;AAED,iGAAiG;AACjG,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,wBAAgB,UAAU,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE,EAAE,IAAI,GAAE,cAAmB,GAAG,YAAY,CAiBhG"}

package/dist/ml/classical/robust_z.js

@@ -0,0 +1,47 @@
+import { featureDim, median } from "./math.js";
+/** MAD → standard deviation for normally-distributed data. */
+const MAD_TO_SD = 1.4826;
+/** Mean absolute deviation → standard deviation for normally-distributed data. */
+const MEANAD_TO_SD = 1.2533;
+export class RobustZModel {
+    center;
+    scale;
+    threshold;
+    constructor(center, scale, threshold) {
+        this.center = center;
+        this.scale = scale;
+        this.threshold = threshold;
+    }
+    /** Per-feature robust z-scores (absolute). */
+    zScores(x) {
+        return Array.from(this.center, (c, i) => Math.abs((x[i] ?? 0) - c) / this.scale[i]);
+    }
+    predict(x) {
+        const score = Math.max(0, ...this.zScores(x)) / this.threshold;
+        return { scores: [score], confidence: anomalyConfidence(score) };
+    }
+}
+/** Confidence of an anomaly score normalised so 1 is the threshold: distance from it, capped. */
+export function anomalyConfidence(score) {
+    return Math.max(0, Math.min(1, Math.abs(score - 1)));
+}
+export function fitRobustZ(samples, opts = {}) {
+    const X = samples.map((s) => s.x);
+    const d = featureDim(X);
+    const minFrac = opts.minScaleFraction ?? 0.01;
+    const minAbs = opts.minScale ?? 0;
+    const center = new Float64Array(d);
+    const scale = new Float64Array(d);
+    for (let i = 0; i < d; i++) {
+        const col = X.map((x) => x[i]);
+        const med = median(col);
+        const dev = col.map((v) => Math.abs(v - med));
+        let s = MAD_TO_SD * median(dev);
+        if (!(s > 0))
+            s = MEANAD_TO_SD * (dev.reduce((a, b) => a + b, 0) / dev.length);
+        center[i] = med;
+        scale[i] = Math.max(s, minFrac * Math.abs(med) + 1e-9, minAbs);
+    }
+    return new RobustZModel(center, scale, opts.threshold ?? 3.5);
+}
+//# sourceMappingURL=robust_z.js.map

package/dist/ml/classical/robust_z.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"robust_z.js","sourceRoot":"","sources":["../../../src/ml/classical/robust_z.ts"],"names":[],"mappings":"AAcA,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,MAAM,WAAW,CAAC;AAE/C,8DAA8D;AAC9D,MAAM,SAAS,GAAG,MAAM,CAAC;AACzB,kFAAkF;AAClF,MAAM,YAAY,GAAG,MAAM,CAAC;AAc5B,MAAM,OAAO,YAAY;IAEZ;IACA;IACA;IAHX,YACW,MAAoB,EACpB,KAAmB,EACnB,SAAiB;QAFjB,WAAM,GAAN,MAAM,CAAc;QACpB,UAAK,GAAL,KAAK,CAAc;QACnB,cAAS,GAAT,SAAS,CAAQ;IACzB,CAAC;IAEJ,8CAA8C;IAC9C,OAAO,CAAC,CAAoB;QAC1B,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC;IACvF,CAAC;IAED,OAAO,CAAC,CAAe;QACrB,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,GAAG,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC;QAC/D,OAAO,EAAE,MAAM,EAAE,CAAC,KAAK,CAAC,EAAE,UAAU,EAAE,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;IACnE,CAAC;CACF;AAED,iGAAiG;AACjG,MAAM,UAAU,iBAAiB,CAAC,KAAa;IAC7C,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,UAAU,UAAU,CAAC,OAA4B,EAAE,OAAuB,EAAE;IAChF,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAClC,MAAM,CAAC,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC;IACxB,MAAM,OAAO,GAAG,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAAC;IAC9C,MAAM,MAAM,GAAG,IAAI,CAAC,QAAQ,IAAI,CAAC,CAAC;IAClC,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IACnC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3B,MAAM,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;QAChC,MAAM,GAAG,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;QACxB,MAAM,GAAG,GAAG,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC;QAC9C,IAAI,CAAC,GAAG,SAAS,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;QAChC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;YAAE,CAAC,GAAG,YAAY,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC;QAC/E,MAAM,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;QAChB,KAAK,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,IAAI,EAAE,MAAM,CAAC,CAAC;IACjE,CAAC;IACD,OAAO,IAAI,YAAY,CAAC,MAAM,EAAE,KAAK,EAAE,IAAI,CAAC,SAAS,IAAI,GAAG,CAAC,CAAC;AAChE,CAAC"}

package/dist/ml/dense_head.d.ts

@@ -0,0 +1,33 @@
+import { type HeadParam, type TrainableHead } from "./head_trainer.js";
+import type { MlPrediction, MlPredictor, MlSample } from "./types.js";
+export interface DenseHeadConfig {
+    inputDim: number;
+    /** Hidden width; 0 = a linear head. */
+    hiddenDim: number;
+    /** Classes (softmax) or 1 (mse regression). */
+    outputDim: number;
+    /** `softmax` = classification over `outputDim` classes; `mse` = scalar regression. */
+    objective: "softmax" | "mse";
+    seed?: number;
+}
+export declare class DenseHead implements TrainableHead<MlSample>, MlPredictor {
+    readonly config: Required<DenseHeadConfig>;
+    private readonly _params;
+    private readonly _grads;
+    constructor(config: DenseHeadConfig);
+    private _p;
+    private _g;
+    parameters(): HeadParam[];
+    gradients(): HeadParam[];
+    zeroGrad(): void;
+    /** Hidden activation (or the input itself for a linear head) and the output logits/value. */
+    private _forward;
+    /** Loss and dLoss/dz of one sample (unweighted). */
+    private _lossGrad;
+    loss(s: MlSample): number;
+    accumulate(s: MlSample): number;
+    predict(x: Float32Array): MlPrediction;
+    exportParams(): Float32Array;
+    importParams(params: ArrayLike<number>): void;
+}
+//# sourceMappingURL=dense_head.d.ts.map

package/dist/ml/dense_head.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"dense_head.d.ts","sourceRoot":"","sources":["../../src/ml/dense_head.ts"],"names":[],"mappings":"AAUA,OAAO,EAA+B,KAAK,SAAS,EAAE,KAAK,aAAa,EAAE,MAAM,mBAAmB,CAAC;AACpG,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAEtE,MAAM,WAAW,eAAe;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,uCAAuC;IACvC,SAAS,EAAE,MAAM,CAAC;IAClB,+CAA+C;IAC/C,SAAS,EAAE,MAAM,CAAC;IAClB,sFAAsF;IACtF,SAAS,EAAE,SAAS,GAAG,KAAK,CAAC;IAC7B,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,qBAAa,SAAU,YAAW,aAAa,CAAC,QAAQ,CAAC,EAAE,WAAW;IACpE,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,eAAe,CAAC,CAAC;IAC3C,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAc;IACtC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAc;gBAEzB,MAAM,EAAE,eAAe;IAsBnC,OAAO,CAAC,EAAE;IAIV,OAAO,CAAC,EAAE;IAIV,UAAU,IAAI,SAAS,EAAE;IAIzB,SAAS,IAAI,SAAS,EAAE;IAIxB,QAAQ,IAAI,IAAI;IAIhB,6FAA6F;IAC7F,OAAO,CAAC,QAAQ;IA0BhB,oDAAoD;IACpD,OAAO,CAAC,SAAS;IAUjB,IAAI,CAAC,CAAC,EAAE,QAAQ,GAAG,MAAM;IAIzB,UAAU,CAAC,CAAC,EAAE,QAAQ,GAAG,MAAM;IA6B/B,OAAO,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY;IAOtC,YAAY,IAAI,YAAY;IAI5B,YAAY,CAAC,MAAM,EAAE,SAAS,CAAC,MAAM,CAAC,GAAG,IAAI;CAG9C"}

package/dist/ml/dense_head.js

@@ -0,0 +1,140 @@
+/**
+ * ml/dense_head.ts — the catalog's trainable head: linear, or one tanh hidden layer,
+ * over a feature vector (usually an experience embedding).
+ *
+ * It is a {@link TrainableHead} (so the ONE {@link HeadTrainer} fits it, on the
+ * GPU when there is one) and an {@link MlPredictor} (so the runner answers with
+ * it). Its parameters flatten to one vector — the unit that federates.
+ */
+import { gaussianArray, SeededRng } from "../utils/rng.js";
+import { softmax } from "../training/cpu_ops.js";
+import { assignParams, flattenParams } from "./head_trainer.js";
+export class DenseHead {
+    config;
+    _params;
+    _grads;
+    constructor(config) {
+        this.config = { seed: 0x4e4d, ...config };
+        const { inputDim, hiddenDim, outputDim, objective } = this.config;
+        if (objective === "mse" && outputDim !== 1)
+            throw new Error("DenseHead: mse objective needs outputDim 1");
+        if (objective === "softmax" && outputDim < 2)
+            throw new Error("DenseHead: softmax objective needs outputDim ≥ 2");
+        const rng = new SeededRng(this.config.seed);
+        const next = () => rng.next();
+        const inner = hiddenDim > 0 ? hiddenDim : inputDim;
+        const shapes = [
+            ...(hiddenDim > 0
+                ? [
+                    ["w1", hiddenDim * inputDim, 1 / Math.sqrt(inputDim)],
+                    ["b1", hiddenDim, 0],
+                ]
+                : []),
+            ["w2", outputDim * inner, 1 / Math.sqrt(inner)],
+            ["b2", outputDim, 0],
+        ];
+        this._params = shapes.map(([name, n, std]) => ({ name, numel: n, data: std > 0 ? gaussianArray(n, std, next) : new Float32Array(n) }));
+        this._grads = shapes.map(([name, n]) => ({ name, numel: n, data: new Float32Array(n) }));
+    }
+    _p(name) {
+        return this._params.find((p) => p.name === name).data;
+    }
+    _g(name) {
+        return this._grads.find((p) => p.name === name).data;
+    }
+    parameters() {
+        return this._params;
+    }
+    gradients() {
+        return this._grads;
+    }
+    zeroGrad() {
+        for (const g of this._grads)
+            g.data.fill(0);
+    }
+    /** Hidden activation (or the input itself for a linear head) and the output logits/value. */
+    _forward(x) {
+        const { inputDim, hiddenDim, outputDim } = this.config;
+        let h;
+        if (hiddenDim > 0) {
+            const w1 = this._p("w1");
+            const b1 = this._p("b1");
+            h = new Float32Array(hiddenDim);
+            for (let j = 0; j < hiddenDim; j++) {
+                let s = b1[j];
+                for (let i = 0; i < inputDim; i++)
+                    s += w1[j * inputDim + i] * (x[i] ?? 0);
+                h[j] = Math.tanh(s);
+            }
+        }
+        else {
+            h = Float32Array.from({ length: inputDim }, (_, i) => x[i] ?? 0);
+        }
+        const w2 = this._p("w2");
+        const b2 = this._p("b2");
+        const z = new Float32Array(outputDim);
+        for (let k = 0; k < outputDim; k++) {
+            let s = b2[k];
+            for (let j = 0; j < h.length; j++)
+                s += w2[k * h.length + j] * h[j];
+            z[k] = s;
+        }
+        return { h, z };
+    }
+    /** Loss and dLoss/dz of one sample (unweighted). */
+    _lossGrad(z, y) {
+        if (this.config.objective === "mse") {
+            const d = z[0] - y;
+            return { loss: 0.5 * d * d, dz: Float32Array.of(d) };
+        }
+        const p = softmax(z);
+        const dz = Float32Array.from(p, (v, k) => v - (k === y ? 1 : 0));
+        return { loss: -Math.log(Math.max(p[y] ?? 0, 1e-12)), dz };
+    }
+    loss(s) {
+        return (s.weight ?? 1) * this._lossGrad(this._forward(s.x).z, s.y).loss;
+    }
+    accumulate(s) {
+        const w = s.weight ?? 1;
+        const { inputDim, hiddenDim, outputDim } = this.config;
+        const { h, z } = this._forward(s.x);
+        const { loss, dz } = this._lossGrad(z, s.y);
+        const w2 = this._p("w2");
+        const gW2 = this._g("w2");
+        const gB2 = this._g("b2");
+        const dh = new Float32Array(h.length);
+        for (let k = 0; k < outputDim; k++) {
+            const g = w * dz[k];
+            gB2[k] = gB2[k] + g;
+            for (let j = 0; j < h.length; j++) {
+                gW2[k * h.length + j] = gW2[k * h.length + j] + g * h[j];
+                dh[j] = dh[j] + g * w2[k * h.length + j];
+            }
+        }
+        if (hiddenDim > 0) {
+            const gW1 = this._g("w1");
+            const gB1 = this._g("b1");
+            for (let j = 0; j < hiddenDim; j++) {
+                const dpre = dh[j] * (1 - h[j] * h[j]);
+                gB1[j] = gB1[j] + dpre;
+                for (let i = 0; i < inputDim; i++)
+                    gW1[j * inputDim + i] = gW1[j * inputDim + i] + dpre * (s.x[i] ?? 0);
+            }
+        }
+        return w * loss;
+    }
+    predict(x) {
+        const { z } = this._forward(x);
+        if (this.config.objective === "mse")
+            return { scores: [z[0]], confidence: 1 };
+        const scores = Array.from(softmax(z));
+        return { scores, confidence: Math.max(...scores) };
+    }
+    exportParams() {
+        return flattenParams(this._params);
+    }
+    importParams(params) {
+        assignParams(this._params, params);
+    }
+}
+//# sourceMappingURL=dense_head.js.map

package/dist/ml/dense_head.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"dense_head.js","sourceRoot":"","sources":["../../src/ml/dense_head.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,aAAa,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAC3D,OAAO,EAAE,OAAO,EAAE,MAAM,wBAAwB,CAAC;AACjD,OAAO,EAAE,YAAY,EAAE,aAAa,EAAsC,MAAM,mBAAmB,CAAC;AAcpG,MAAM,OAAO,SAAS;IACX,MAAM,CAA4B;IAC1B,OAAO,CAAc;IACrB,MAAM,CAAc;IAErC,YAAY,MAAuB;QACjC,IAAI,CAAC,MAAM,GAAG,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,MAAM,EAAE,CAAC;QAC1C,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QAClE,IAAI,SAAS,KAAK,KAAK,IAAI,SAAS,KAAK,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,4CAA4C,CAAC,CAAC;QAC1G,IAAI,SAAS,KAAK,SAAS,IAAI,SAAS,GAAG,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,kDAAkD,CAAC,CAAC;QAClH,MAAM,GAAG,GAAG,IAAI,SAAS,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;QAC5C,MAAM,IAAI,GAAG,GAAG,EAAE,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC;QAC9B,MAAM,KAAK,GAAG,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,QAAQ,CAAC;QACnD,MAAM,MAAM,GAA+B;YACzC,GAAG,CAAC,SAAS,GAAG,CAAC;gBACf,CAAC,CAAE;oBACC,CAAC,IAAI,EAAE,SAAS,GAAG,QAAQ,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;oBACrD,CAAC,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;iBACU;gBAClC,CAAC,CAAC,EAAE,CAAC;YACP,CAAC,IAAI,EAAE,SAAS,GAAG,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAC/C,CAAC,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;SACrB,CAAC;QACF,IAAI,CAAC,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,EAAE,GAAG,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACvI,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,IAAI,EAAE,IAAI,YAAY,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAC3F,CAAC;IAEO,EAAE,CAAC,IAAY;QACrB,OAAO,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAE,CAAC,IAAI,CAAC;IACzD,CAAC;IAEO,EAAE,CAAC,IAAY;QACrB,OAAO,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAE,CAAC,IAAI,CAAC;IACxD,CAAC;IAED,UAAU;QACR,OAAO,IAAI,CAAC,OAAO,CAAC;IACtB,CAAC;IAED,SAAS;QACP,OAAO,IAAI,CAAC,MAAM,CAAC;IACrB,CAAC;IAED,QAAQ;QACN,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,MAAM;YAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC9C,CAAC;IAED,6FAA6F;IACrF,QAAQ,CAAC,CAAoB;QACnC,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QACvD,IAAI,CAAe,CAAC;QACpB,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,EAAE,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;YACzB,MAAM,EAAE,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;YACzB,CAAC,GAAG,IAAI,YAAY,CAAC,SAAS,CAAC,CAAC;YAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;gBACnC,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,CAAC;gBACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,EAAE,CAAC,EAAE;oBAAE,CAAC,IAAI,EAAE,CAAC,CAAC,GAAG,QAAQ,GAAG,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;gBAC5E,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YACtB,CAAC;QACH,CAAC;aAAM,CAAC;YACN,CAAC,GAAG,YAAY,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;QACnE,CAAC;QACD,MAAM,EAAE,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;QACzB,MAAM,EAAE,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;QACzB,MAAM,CAAC,GAAG,IAAI,YAAY,CAAC,SAAS,CAAC,CAAC;QACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;YACnC,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,CAAC;YACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;gBAAE,CAAC,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;YACtE,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;QACX,CAAC;QACD,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC;IAClB,CAAC;IAED,oDAAoD;IAC5C,SAAS,CAAC,CAAe,EAAE,CAAS;QAC1C,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS,KAAK,KAAK,EAAE,CAAC;YACpC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC;YACpB,OAAO,EAAE,IAAI,EAAE,GAAG,GAAG,CAAC,GAAG,CAAC,EAAE,EAAE,EAAE,YAAY,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvD,CAAC;QACD,MAAM,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACrB,MAAM,EAAE,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACjE,OAAO,EAAE,IAAI,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC;IAC7D,CAAC;IAED,IAAI,CAAC,CAAW;QACd,OAAO,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;IAC1E,CAAC;IAED,UAAU,CAAC,CAAW;QACpB,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC;QACxB,MAAM,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC;QACvD,MAAM,EAAE,CAAC,EAAE,CAAC,EAAE,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACpC,MAAM,EAAE,IAAI,EAAE,EAAE,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAC5C,MAAM,EAAE,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;QACzB,MAAM,GAAG,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;QAC1B,MAAM,GAAG,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;QAC1B,MAAM,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;YACnC,MAAM,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,CAAC;YACrB,GAAG,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC;YACrB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;gBAC3D,EAAE,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,CAAE,GAAG,CAAC,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAE,CAAC;YAC7C,CAAC;QACH,CAAC;QACD,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,GAAG,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;YAC1B,MAAM,GAAG,GAAG,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;YAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;gBACnC,MAAM,IAAI,GAAG,EAAE,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;gBAC1C,GAAG,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAE,GAAG,IAAI,CAAC;gBACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,EAAE,CAAC,EAAE;oBAAE,GAAG,CAAC,CAAC,GAAG,QAAQ,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,QAAQ,GAAG,CAAC,CAAE,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;YAC3G,CAAC;QACH,CAAC;QACD,OAAO,CAAC,GAAG,IAAI,CAAC;IAClB,CAAC;IAED,OAAO,CAAC,CAAe;QACrB,MAAM,EAAE,CAAC,EAAE,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;QAC/B,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS,KAAK,KAAK;YAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,CAAC;QAC/E,MAAM,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QACtC,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC;IACrD,CAAC;IAED,YAAY;QACV,OAAO,aAAa,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;IACrC,CAAC;IAED,YAAY,CAAC,MAAyB;QACpC,YAAY,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;IACrC,CAAC;CACF"}

package/dist/ml/eval_loss.d.ts

@@ -0,0 +1,15 @@
+/**
+ * ml/eval_loss.ts — the ONE held-out loss for catalog models.
+ *
+ * What the trainer reports as `holdoutLoss`, what a federated merge is gated on, and
+ * what `LogisticModel.loss` returns. A leaf (types only), so a classical algorithm can
+ * use it without importing the trainer that imports the algorithms.
+ */
+import type { MlPredictor, MlSample, MlTaskKind } from "./types.js";
+/**
+ * Mean loss of `predictor` on `samples` for a task of `kind` — classify: weighted
+ * cross-entropy of the true class; regress: weighted ½·squared error. Null for
+ * kinds without a supervised loss, or an empty (or zero-weight) set.
+ */
+export declare function mlEvalLoss(kind: MlTaskKind, predictor: MlPredictor, samples: readonly MlSample[]): number | null;
+//# sourceMappingURL=eval_loss.d.ts.map

package/dist/ml/eval_loss.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval_loss.d.ts","sourceRoot":"","sources":["../../src/ml/eval_loss.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,OAAO,KAAK,EAAE,WAAW,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,YAAY,CAAC;AAEpE;;;;GAIG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,WAAW,EAAE,OAAO,EAAE,SAAS,QAAQ,EAAE,GAAG,MAAM,GAAG,IAAI,CAYhH"}

package/dist/ml/eval_loss.js

@@ -0,0 +1,20 @@
+/**
+ * Mean loss of `predictor` on `samples` for a task of `kind` — classify: weighted
+ * cross-entropy of the true class; regress: weighted ½·squared error. Null for
+ * kinds without a supervised loss, or an empty (or zero-weight) set.
+ */
+export function mlEvalLoss(kind, predictor, samples) {
+    if (samples.length === 0 || (kind !== "classify" && kind !== "regress"))
+        return null;
+    let sum = 0;
+    let wSum = 0;
+    for (const s of samples) {
+        const w = s.weight ?? 1;
+        const p = predictor.predict(s.x);
+        const loss = kind === "classify" ? -Math.log(Math.max(p.scores[s.y] ?? 0, 1e-12)) : 0.5 * (p.scores[0] - s.y) ** 2;
+        sum += w * loss;
+        wSum += w;
+    }
+    return wSum > 0 ? sum / wSum : null;
+}
+//# sourceMappingURL=eval_loss.js.map

package/dist/ml/eval_loss.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval_loss.js","sourceRoot":"","sources":["../../src/ml/eval_loss.ts"],"names":[],"mappings":"AASA;;;;GAIG;AACH,MAAM,UAAU,UAAU,CAAC,IAAgB,EAAE,SAAsB,EAAE,OAA4B;IAC/F,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,IAAI,KAAK,UAAU,IAAI,IAAI,KAAK,SAAS,CAAC;QAAE,OAAO,IAAI,CAAC;IACrF,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC;QACxB,MAAM,CAAC,GAAG,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACjC,MAAM,IAAI,GAAG,IAAI,KAAK,UAAU,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QACpH,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC;QAChB,IAAI,IAAI,CAAC,CAAC;IACZ,CAAC;IACD,OAAO,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC;AACtC,CAAC"}

package/dist/ml/features.d.ts

@@ -0,0 +1,25 @@
+/**
+ * ml/features.ts — how an input becomes a feature vector.
+ *
+ * Text goes through the existing CPU embedder seam ({@link TextEmbedder}, which
+ * `EvermindTextEmbedder` implements: the person's own Evermind turns text into
+ * an L2-normalised vector with no GPU and no network). Numeric fields are
+ * appended as the task supplies them — on a FIXED scale (see {@link logScale}),
+ * never one fitted to local data, so a federated weight means the same thing on
+ * every machine.
+ */
+import type { TextEmbedder } from "../lm/text_embedder.js";
+import type { FeatureSpec } from "./types.js";
+export type { TextEmbedder } from "../lm/text_embedder.js";
+/** Feature vector of one input: `[embed(text) …, numeric …]`. */
+export declare function extractFeatures<I>(spec: FeatureSpec<I>, input: I, embedder?: TextEmbedder): Promise<Float32Array>;
+/** Feature vectors of many inputs, in order. */
+export declare function extractFeatureMatrix<I>(spec: FeatureSpec<I>, inputs: readonly I[], embedder?: TextEmbedder): Promise<Float32Array[]>;
+/**
+ * A count or size on a fixed log scale: log1p(max(v, 0)) / log1p(ref). `ref` is
+ * the value that maps to 1 (e.g. 100k tokens), so the scale is the same everywhere.
+ */
+export declare function logScale(v: number, ref: number): number;
+/** A ratio clamped to [0, max] — fill fractions, success rates. */
+export declare function clampRatio(numerator: number, denominator: number, max?: number): number;
+//# sourceMappingURL=features.d.ts.map

package/dist/ml/features.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"features.d.ts","sourceRoot":"","sources":["../../src/ml/features.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,wBAAwB,CAAC;AAC3D,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAE9C,YAAY,EAAE,YAAY,EAAE,MAAM,wBAAwB,CAAC;AAE3D,iEAAiE;AACjE,wBAAsB,eAAe,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,QAAQ,CAAC,EAAE,YAAY,GAAG,OAAO,CAAC,YAAY,CAAC,CAmBvH;AAED,gDAAgD;AAChD,wBAAsB,oBAAoB,CAAC,CAAC,EAC1C,IAAI,EAAE,WAAW,CAAC,CAAC,CAAC,EACpB,MAAM,EAAE,SAAS,CAAC,EAAE,EACpB,QAAQ,CAAC,EAAE,YAAY,GACtB,OAAO,CAAC,YAAY,EAAE,CAAC,CAIzB;AAED;;;GAGG;AACH,wBAAgB,QAAQ,CAAC,CAAC,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED,mEAAmE;AACnE,wBAAgB,UAAU,CAAC,SAAS,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,GAAG,SAAI,GAAG,MAAM,CAGlF"}

package/dist/ml/features.js

@@ -0,0 +1,46 @@
+/** Feature vector of one input: `[embed(text) …, numeric …]`. */
+export async function extractFeatures(spec, input, embedder) {
+    const parts = [];
+    if (spec.text) {
+        if (!embedder)
+            throw new Error("extractFeatures: this task embeds text but no embedder was supplied");
+        parts.push(await embedder.embed(spec.text(input)));
+    }
+    if (spec.numeric) {
+        const nums = spec.numeric(input);
+        for (const v of nums)
+            if (!Number.isFinite(v))
+                throw new Error(`extractFeatures: non-finite numeric feature ${v}`);
+        parts.push(nums);
+    }
+    if (parts.length === 0)
+        throw new Error("extractFeatures: the feature spec has neither text nor numeric fields");
+    const out = new Float32Array(parts.reduce((n, p) => n + p.length, 0));
+    let o = 0;
+    for (const p of parts) {
+        out.set(p, o);
+        o += p.length;
+    }
+    return out;
+}
+/** Feature vectors of many inputs, in order. */
+export async function extractFeatureMatrix(spec, inputs, embedder) {
+    const out = [];
+    for (const i of inputs)
+        out.push(await extractFeatures(spec, i, embedder));
+    return out;
+}
+/**
+ * A count or size on a fixed log scale: log1p(max(v, 0)) / log1p(ref). `ref` is
+ * the value that maps to 1 (e.g. 100k tokens), so the scale is the same everywhere.
+ */
+export function logScale(v, ref) {
+    return Math.log1p(Math.max(0, v)) / Math.log1p(Math.max(ref, 1));
+}
+/** A ratio clamped to [0, max] — fill fractions, success rates. */
+export function clampRatio(numerator, denominator, max = 4) {
+    if (!(denominator > 0))
+        return 0;
+    return Math.max(0, Math.min(max, numerator / denominator));
+}
+//# sourceMappingURL=features.js.map

package/dist/ml/features.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"features.js","sourceRoot":"","sources":["../../src/ml/features.ts"],"names":[],"mappings":"AAeA,iEAAiE;AACjE,MAAM,CAAC,KAAK,UAAU,eAAe,CAAI,IAAoB,EAAE,KAAQ,EAAE,QAAuB;IAC9F,MAAM,KAAK,GAAwB,EAAE,CAAC;IACtC,IAAI,IAAI,CAAC,IAAI,EAAE,CAAC;QACd,IAAI,CAAC,QAAQ;YAAE,MAAM,IAAI,KAAK,CAAC,qEAAqE,CAAC,CAAC;QACtG,KAAK,CAAC,IAAI,CAAC,MAAM,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACrD,CAAC;IACD,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;QACjB,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;QACjC,KAAK,MAAM,CAAC,IAAI,IAAI;YAAE,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC;gBAAE,MAAM,IAAI,KAAK,CAAC,+CAA+C,CAAC,EAAE,CAAC,CAAC;QACnH,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACnB,CAAC;IACD,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,uEAAuE,CAAC,CAAC;IACjH,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC;IACtE,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACtB,GAAG,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACd,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC;IAChB,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,gDAAgD;AAChD,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,IAAoB,EACpB,MAAoB,EACpB,QAAuB;IAEvB,MAAM,GAAG,GAAmB,EAAE,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,MAAM;QAAE,GAAG,CAAC,IAAI,CAAC,MAAM,eAAe,CAAC,IAAI,EAAE,CAAC,EAAE,QAAQ,CAAC,CAAC,CAAC;IAC3E,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,QAAQ,CAAC,CAAS,EAAE,GAAW;IAC7C,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC;AACnE,CAAC;AAED,mEAAmE;AACnE,MAAM,UAAU,UAAU,CAAC,SAAiB,EAAE,WAAmB,EAAE,GAAG,GAAG,CAAC;IACxE,IAAI,CAAC,CAAC,WAAW,GAAG,CAAC,CAAC;QAAE,OAAO,CAAC,CAAC;IACjC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,SAAS,GAAG,WAAW,CAAC,CAAC,CAAC;AAC7D,CAAC"}