@seanhogg/builderforce-memory-engine 2026.9.33 → 2026.10.1

package/dist/limbic/limbic_task.d.ts

@@ -0,0 +1,22 @@
+import type { MlPredictor, MlTask } from "../ml/types.js";
+import type { LimbicModel } from "./limbic_model.js";
+export interface LimbicAffectInput {
+    /** Experience embedding (length = model.inputDim). */
+    embedding: ArrayLike<number>;
+    /** Current affective state (length = model.stateDim). */
+    state: ArrayLike<number>;
+}
+export interface LimbicAffectOutput {
+    /** Bounded affect delta in (-1, 1) per state dim. */
+    delta: Float32Array;
+    /** Reward prediction. */
+    reward: number;
+}
+export declare const LIMBIC_AFFECT_TASK: MlTask<LimbicAffectInput, LimbicAffectOutput>;
+/**
+ * Answer `limbic.affect` with `model`. The feature vector is `[embedding, state]`;
+ * the hidden scratch resets per call, exactly as the trainer treats each sample.
+ * Confidence is 1: a deterministic regression head has no calibrated uncertainty.
+ */
+export declare function limbicAffectPredictor(model: LimbicModel): MlPredictor;
+//# sourceMappingURL=limbic_task.d.ts.map

package/dist/limbic/limbic_task.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"limbic_task.d.ts","sourceRoot":"","sources":["../../src/limbic/limbic_task.ts"],"names":[],"mappings":"AAWA,OAAO,KAAK,EAAgB,WAAW,EAAE,MAAM,EAAE,MAAM,gBAAgB,CAAC;AACxE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAErD,MAAM,WAAW,iBAAiB;IAChC,sDAAsD;IACtD,SAAS,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IAC7B,yDAAyD;IACzD,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;CAC1B;AAED,MAAM,WAAW,kBAAkB;IACjC,qDAAqD;IACrD,KAAK,EAAE,YAAY,CAAC;IACpB,yBAAyB;IACzB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,eAAO,MAAM,kBAAkB,EAAE,MAAM,CAAC,iBAAiB,EAAE,kBAAkB,CAa5E,CAAC;AAEF;;;;GAIG;AACH,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,WAAW,GAAG,WAAW,CAarE"}

package/dist/limbic/limbic_task.js

@@ -0,0 +1,45 @@
+/**
+ * limbic/limbic_task.ts — `limbic.affect`, the ML catalog's first row, registered
+ * by limbic/ itself.
+ *
+ * Nothing about the affect head changes: the row describes it (input = experience
+ * embedding + affective state, output = bounded affect delta + reward), the
+ * predictor adapter lets the ONE runner answer with an existing `LimbicModel`, and
+ * training stays `LimbicTrainer` (itself a thin registration over `HeadTrainer`).
+ * Its flat parameters are what federate when the head travels with the model.
+ */
+import { assignParams, flattenParams } from "../ml/head_trainer.js";
+export const LIMBIC_AFFECT_TASK = {
+    kind: "regress",
+    id: "limbic.affect",
+    description: "How an experience moves the agent's affective state, and how rewarding it was.",
+    features: { numeric: (i) => [...Array.from(i.embedding), ...Array.from(i.state)] },
+    engines: [{ tier: "local", engine: "head", head: "mlp" }],
+    train: { minSamples: 1, trainedBy: "LimbicTrainer" },
+    output: {
+        decode: (p) => ({
+            delta: Float32Array.from(p.scores.slice(0, -1)),
+            reward: p.scores[p.scores.length - 1] ?? 0,
+        }),
+    },
+};
+/**
+ * Answer `limbic.affect` with `model`. The feature vector is `[embedding, state]`;
+ * the hidden scratch resets per call, exactly as the trainer treats each sample.
+ * Confidence is 1: a deterministic regression head has no calibrated uncertainty.
+ */
+export function limbicAffectPredictor(model) {
+    const { inputDim, stateDim } = model.config;
+    return {
+        predict(x) {
+            if (x.length !== inputDim + stateDim) {
+                throw new Error(`limbic.affect: expected ${inputDim + stateDim} features (embedding ${inputDim} + state ${stateDim}), got ${x.length}`);
+            }
+            const f = model.forward(x.subarray(0, inputDim), model.initHidden(), x.subarray(inputDim));
+            return { scores: [...Array.from(f.delta), f.reward], confidence: 1 };
+        },
+        exportParams: () => flattenParams(model.parameters()),
+        importParams: (params) => assignParams(model.parameters(), params),
+    };
+}
+//# sourceMappingURL=limbic_task.js.map

package/dist/limbic/limbic_task.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"limbic_task.js","sourceRoot":"","sources":["../../src/limbic/limbic_task.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AAkBpE,MAAM,CAAC,MAAM,kBAAkB,GAAkD;IAC/E,IAAI,EAAE,SAAS;IACf,EAAE,EAAE,eAAe;IACnB,WAAW,EAAE,gFAAgF;IAC7F,QAAQ,EAAE,EAAE,OAAO,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE;IAClF,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC;IACzD,KAAK,EAAE,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,eAAe,EAAE;IACpD,MAAM,EAAE;QACN,MAAM,EAAE,CAAC,CAAe,EAAE,EAAE,CAAC,CAAC;YAC5B,KAAK,EAAE,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;YAC/C,MAAM,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,IAAI,CAAC;SAC3C,CAAC;KACH;CACF,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CAAC,KAAkB;IACtD,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;IAC5C,OAAO;QACL,OAAO,CAAC,CAAe;YACrB,IAAI,CAAC,CAAC,MAAM,KAAK,QAAQ,GAAG,QAAQ,EAAE,CAAC;gBACrC,MAAM,IAAI,KAAK,CAAC,2BAA2B,QAAQ,GAAG,QAAQ,wBAAwB,QAAQ,YAAY,QAAQ,UAAU,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;YAC1I,CAAC;YACD,MAAM,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,QAAQ,CAAC,EAAE,KAAK,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC;YAC3F,OAAO,EAAE,MAAM,EAAE,CAAC,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,MAAM,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,CAAC;QACvE,CAAC;QACD,YAAY,EAAE,GAAG,EAAE,CAAC,aAAa,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC;QACrD,YAAY,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,YAAY,CAAC,KAAK,CAAC,UAAU,EAAE,EAAE,MAAM,CAAC;KACnE,CAAC;AACJ,CAAC"}

package/dist/limbic/limbic_trainer.d.ts

@@ -3,15 +3,16 @@
  *
  * Trains the affect head to predict affect deltas and reward from
  * (experience embedding, current state) pairs, using full-batch gradient
- * descent with AdamW. When a GPUDevice is supplied the optimiser step runs on
- * the GPU via the shared WEIGHT_UPDATE_WGSL kernel (real WebGPU training); with
- * no device it uses a numerically-equivalent CPU AdamW so training works
- * everywhere (CI, Node without @webgpu/node, etc.).
+ * descent with AdamW — through the shared {@link HeadTrainer} (`ml/head_trainer.ts`),
+ * which runs the optimiser step on the GPU via WEIGHT_UPDATE_WGSL when a device is
+ * supplied and on a numerically-equivalent CPU AdamW otherwise. This file is only
+ * the limbic registration: the head adapter and the public trainer surface.
  *
  * The objective is MSE (regression), not the cross-entropy used by the
  * language-model {@link MambaTrainer} — a limbic experience has continuous
  * targets, not a next-token distribution.
  */
+import { type HeadTrainOptions, type TrainableHead } from "../ml/head_trainer.js";
 import type { LimbicModel } from "./limbic_model.js";
 /** One training example: an experience and the affect change it should produce. */
 export interface LimbicSample {
@@ -24,43 +25,29 @@ export interface LimbicSample {
     /** Observed scalar reward for the experience. */
     reward: number;
 }
-export interface LimbicTrainOptions {
-    learningRate?: number;
-    epochs?: number;
-    weightDecay?: number;
-    beta1?: number;
-    beta2?: number;
-    eps?: number;
-    /** Max global gradient L2 norm before the optimiser step. Default 1.0. */
-    maxGradNorm?: number;
-    /** Per-epoch hook. `gradNorm` is the pre-clip gradient L2 norm this epoch (the
-     *  instability early-warning) — computed CPU-side so it's always available, unlike
-     *  the GPU trainer where it's opt-in. */
-    onEpochEnd?: ((epoch: number, loss: number, gradNorm?: number) => void) | null;
-}
+/** The limbic trainer's options are the shared head trainer's. */
+export type LimbicTrainOptions = HeadTrainOptions;
+/**
+ * The limbic model as a {@link TrainableHead}. Each sample is one experience
+ * appraisal: cross-experience memory flows through the affective state `s` (fed
+ * back by the runtime), not through the hidden scratch `h`, so the hidden state
+ * resets per sample. This also makes BPTT(1) exact — there is no truncated carry.
+ */
+export declare function limbicHead(model: LimbicModel): TrainableHead<LimbicSample>;
+/** Thin registration of the limbic head over the shared {@link HeadTrainer}. */
 export declare class LimbicTrainer {
     readonly model: LimbicModel;
     readonly device: GPUDevice | null;
-    private _moments;
-    private _step;
-    private readonly _adamwPipeline;
+    private readonly _trainer;
     constructor(model: LimbicModel, device?: GPUDevice | null);
     /** Whether the optimiser step runs on the GPU. */
     get gpuTraining(): boolean;
-    private _initMoments;
     /**
      * Train on a batch of samples for `epochs` passes. Returns the per-epoch mean
-     * loss (monotonically decreasing on a learnable mapping). Full-batch: grads
-     * accumulate across the whole sequence (recurrent hidden carried, reset per
-     * epoch), are averaged, clipped, then applied once per epoch.
+     * loss (monotonically decreasing on a learnable mapping).
      */
     train(samples: LimbicSample[], opts?: LimbicTrainOptions): Promise<number[]>;
     /** Mean MSE loss over samples (no weight update). Hidden resets per sample. */
     evaluate(samples: LimbicSample[]): number;
-    /** Clip grads to `maxNorm` in place; returns the PRE-clip L2 norm (the signal). */
-    private _clipGradients;
-    private _adamwStepCpu;
-    /** AdamW on the GPU via the shared WEIGHT_UPDATE_WGSL kernel. Awaited per step. */
-    private _adamwStepGpu;
 }
 //# sourceMappingURL=limbic_trainer.d.ts.map

package/dist/limbic/limbic_trainer.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"limbic_trainer.d.ts","sourceRoot":"","sources":["../../src/limbic/limbic_trainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAYH,OAAO,KAAK,EAAE,WAAW,EAAe,MAAM,mBAAmB,CAAC;AAElE,mFAAmF;AACnF,MAAM,WAAW,YAAY;IAC3B,sDAAsD;IACtD,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IACzB,+EAA+E;IAC/E,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IACzB,uFAAuF;IACvF,WAAW,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IAC/B,iDAAiD;IACjD,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,kBAAkB;IACjC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,0EAA0E;IAC1E,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB;;6CAEyC;IACzC,UAAU,CAAC,EAAE,CAAC,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,MAAM,KAAK,IAAI,CAAC,GAAG,IAAI,CAAC;CAChF;AAuBD,qBAAa,aAAa;IACxB,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC;IAC5B,QAAQ,CAAC,MAAM,EAAE,SAAS,GAAG,IAAI,CAAC;IAClC,OAAO,CAAC,QAAQ,CAA6B;IAC7C,OAAO,CAAC,KAAK,CAAK;IAClB,OAAO,CAAC,QAAQ,CAAC,cAAc,CAA4B;gBAE/C,KAAK,EAAE,WAAW,EAAE,MAAM,GAAE,SAAS,GAAG,IAAW;IAQ/D,kDAAkD;IAClD,IAAI,WAAW,IAAI,OAAO,CAEzB;IAED,OAAO,CAAC,YAAY;IAQpB;;;;;OAKG;IACG,KAAK,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,IAAI,GAAE,kBAAuB,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IA6DtF,+EAA+E;IAC/E,QAAQ,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,MAAM;IAkBzC,mFAAmF;IACnF,OAAO,CAAC,cAAc;IAWtB,OAAO,CAAC,aAAa;IAqBrB,mFAAmF;YACrE,aAAa;CAkC5B"}
+{"version":3,"file":"limbic_trainer.d.ts","sourceRoot":"","sources":["../../src/limbic/limbic_trainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAe,KAAK,gBAAgB,EAAE,KAAK,aAAa,EAAE,MAAM,uBAAuB,CAAC;AAC/F,OAAO,KAAK,EAAE,WAAW,EAAe,MAAM,mBAAmB,CAAC;AAElE,mFAAmF;AACnF,MAAM,WAAW,YAAY;IAC3B,sDAAsD;IACtD,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IACzB,+EAA+E;IAC/E,KAAK,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IACzB,uFAAuF;IACvF,WAAW,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;IAC/B,iDAAiD;IACjD,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,kEAAkE;AAClE,MAAM,MAAM,kBAAkB,GAAG,gBAAgB,CAAC;AAElD;;;;;GAKG;AACH,wBAAgB,UAAU,CAAC,KAAK,EAAE,WAAW,GAAG,aAAa,CAAC,YAAY,CAAC,CAoB1E;AAED,gFAAgF;AAChF,qBAAa,aAAa;IACxB,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC;IAC5B,QAAQ,CAAC,MAAM,EAAE,SAAS,GAAG,IAAI,CAAC;IAClC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAA4B;gBAEzC,KAAK,EAAE,WAAW,EAAE,MAAM,GAAE,SAAS,GAAG,IAAW;IAM/D,kDAAkD;IAClD,IAAI,WAAW,IAAI,OAAO,CAEzB;IAED;;;OAGG;IACH,KAAK,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,IAAI,GAAE,kBAAuB,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAIhF,+EAA+E;IAC/E,QAAQ,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,MAAM;CAG1C"}

package/dist/limbic/limbic_trainer.js

@@ -3,172 +3,66 @@
  *
  * Trains the affect head to predict affect deltas and reward from
  * (experience embedding, current state) pairs, using full-batch gradient
- * descent with AdamW. When a GPUDevice is supplied the optimiser step runs on
- * the GPU via the shared WEIGHT_UPDATE_WGSL kernel (real WebGPU training); with
- * no device it uses a numerically-equivalent CPU AdamW so training works
- * everywhere (CI, Node without @webgpu/node, etc.).
+ * descent with AdamW — through the shared {@link HeadTrainer} (`ml/head_trainer.ts`),
+ * which runs the optimiser step on the GPU via WEIGHT_UPDATE_WGSL when a device is
+ * supplied and on a numerically-equivalent CPU AdamW otherwise. This file is only
+ * the limbic registration: the head adapter and the public trainer surface.
  *
  * The objective is MSE (regression), not the cross-entropy used by the
  * language-model {@link MambaTrainer} — a limbic experience has continuous
  * targets, not a next-token distribution.
  */
-import { createUniformBuffer, createStorageBuffer, createComputePipeline, createBindGroup, dispatchKernel, readBuffer, cdiv, } from "../utils/gpu_utils.js";
-import { WEIGHT_UPDATE_WGSL } from "../kernels/weight_update.js";
-function packAdamParams(numElements, lr, beta1, beta2, eps, weightDecay, beta1_t, beta2_t) {
-    const buf = new ArrayBuffer(32);
-    new Uint32Array(buf, 0, 1).set([numElements]);
-    new Float32Array(buf, 4, 7).set([lr, beta1, beta2, eps, weightDecay, beta1_t, beta2_t]);
-    return buf;
+import { HeadTrainer } from "../ml/head_trainer.js";
+/**
+ * The limbic model as a {@link TrainableHead}. Each sample is one experience
+ * appraisal: cross-experience memory flows through the affective state `s` (fed
+ * back by the runtime), not through the hidden scratch `h`, so the hidden state
+ * resets per sample. This also makes BPTT(1) exact — there is no truncated carry.
+ */
+export function limbicHead(model) {
+    return {
+        parameters: () => model.parameters(),
+        gradients: () => model.gradients(),
+        zeroGrad: () => model.zeroGrad(),
+        accumulate: (s) => model.backwardStep(s.input, model.initHidden(), s.state, s.deltaTarget, s.reward).loss,
+        loss: (s) => {
+            const { rewardWeight } = model.config;
+            const f = model.forward(s.input, model.initHidden(), s.state);
+            let loss = 0;
+            for (let k = 0; k < f.delta.length; k++) {
+                const d = f.delta[k] - (s.deltaTarget[k] ?? 0);
+                loss += 0.5 * d * d;
+            }
+            const rd = f.reward - s.reward;
+            loss += 0.5 * rewardWeight * rd * rd;
+            return loss;
+        },
+    };
 }
+/** Thin registration of the limbic head over the shared {@link HeadTrainer}. */
 export class LimbicTrainer {
     model;
     device;
-    _moments = null;
-    _step = 0;
-    _adamwPipeline;
+    _trainer;
     constructor(model, device = null) {
         this.model = model;
         this.device = device;
-        this._adamwPipeline = device
-            ? createComputePipeline(device, WEIGHT_UPDATE_WGSL, "adamw_update")
-            : null;
+        this._trainer = new HeadTrainer(limbicHead(model), device, "LimbicTrainer");
     }
     /** Whether the optimiser step runs on the GPU. */
     get gpuTraining() {
-        return this.device != null && this._adamwPipeline != null;
-    }
-    _initMoments() {
-        if (this._moments)
-            return;
-        this._moments = this.model.parameters().map((p) => ({
-            m: new Float32Array(p.numel),
-            v: new Float32Array(p.numel),
-        }));
+        return this._trainer.gpuTraining;
     }
     /**
      * Train on a batch of samples for `epochs` passes. Returns the per-epoch mean
-     * loss (monotonically decreasing on a learnable mapping). Full-batch: grads
-     * accumulate across the whole sequence (recurrent hidden carried, reset per
-     * epoch), are averaged, clipped, then applied once per epoch.
+     * loss (monotonically decreasing on a learnable mapping).
      */
-    async train(samples, opts = {}) {
-        if (samples.length === 0)
-            throw new Error("LimbicTrainer.train: no samples");
-        const { learningRate = 0.05, epochs = 50, weightDecay = 0.0, beta1 = 0.9, beta2 = 0.999, eps = 1e-8, maxGradNorm = 1.0, onEpochEnd = null, } = opts;
-        this._initMoments();
-        const losses = [];
-        for (let epoch = 0; epoch < epochs; epoch++) {
-            this.model.zeroGrad();
-            let epochLoss = 0;
-            // Each sample is one experience appraisal: cross-experience memory flows
-            // through the affective state `s` (fed back by the runtime), not through
-            // the hidden scratch `h`, so the hidden state resets per sample. This also
-            // makes BPTT(1) exact — there is no truncated carry.
-            for (const s of samples) {
-                const { loss } = this.model.backwardStep(s.input, this.model.initHidden(), s.state, s.deltaTarget, s.reward);
-                epochLoss += loss;
-            }
-            // Average gradients over the batch.
-            const grads = this.model.gradients();
-            const invN = 1 / samples.length;
-            for (const g of grads) {
-                for (let i = 0; i < g.data.length; i++)
-                    g.data[i] *= invN;
-            }
-            const gradNorm = this._clipGradients(grads, maxGradNorm);
-            this._step++;
-            const beta1_t = Math.pow(beta1, this._step);
-            const beta2_t = Math.pow(beta2, this._step);
-            const hp = { learningRate, weightDecay, beta1, beta2, eps, beta1_t, beta2_t };
-            if (this.gpuTraining) {
-                await this._adamwStepGpu(grads, hp);
-            }
-            else {
-                this._adamwStepCpu(grads, hp);
-            }
-            const avg = epochLoss / samples.length;
-            losses.push(avg);
-            if (onEpochEnd)
-                onEpochEnd(epoch + 1, avg, gradNorm);
-        }
-        return losses;
+    train(samples, opts = {}) {
+        return this._trainer.train(samples, opts);
     }
     /** Mean MSE loss over samples (no weight update). Hidden resets per sample. */
     evaluate(samples) {
-        if (samples.length === 0)
-            return 0;
-        let total = 0;
-        const { rewardWeight } = this.model.config;
-        for (const s of samples) {
-            const f = this.model.forward(s.input, this.model.initHidden(), s.state);
-            let loss = 0;
-            for (let k = 0; k < f.delta.length; k++) {
-                const d = f.delta[k] - (s.deltaTarget[k] ?? 0);
-                loss += 0.5 * d * d;
-            }
-            const rd = f.reward - s.reward;
-            loss += 0.5 * rewardWeight * rd * rd;
-            total += loss;
-        }
-        return total / samples.length;
-    }
-    /** Clip grads to `maxNorm` in place; returns the PRE-clip L2 norm (the signal). */
-    _clipGradients(grads, maxNorm) {
-        let normSq = 0;
-        for (const g of grads)
-            for (let i = 0; i < g.data.length; i++)
-                normSq += g.data[i] * g.data[i];
-        const norm = Math.sqrt(normSq);
-        if (norm > maxNorm && norm > 0) {
-            const scale = maxNorm / norm;
-            for (const g of grads)
-                for (let i = 0; i < g.data.length; i++)
-                    g.data[i] *= scale;
-        }
-        return norm;
-    }
-    _adamwStepCpu(grads, hp) {
-        const params = this.model.parameters();
-        const { learningRate: lr, weightDecay: wd, beta1, beta2, eps, beta1_t, beta2_t } = hp;
-        for (let pi = 0; pi < params.length; pi++) {
-            const p = params[pi].data;
-            const g = grads[pi].data;
-            const mom = this._moments[pi];
-            for (let i = 0; i < p.length; i++) {
-                const gi = g[i];
-                mom.m[i] = beta1 * mom.m[i] + (1 - beta1) * gi;
-                mom.v[i] = beta2 * mom.v[i] + (1 - beta2) * gi * gi;
-                const mHat = mom.m[i] / (1 - beta1_t);
-                const vHat = mom.v[i] / (1 - beta2_t);
-                p[i] = p[i] * (1 - lr * wd) - (lr * mHat) / (Math.sqrt(vHat) + eps);
-            }
-        }
-    }
-    /** AdamW on the GPU via the shared WEIGHT_UPDATE_WGSL kernel. Awaited per step. */
-    async _adamwStepGpu(grads, hp) {
-        const device = this.device;
-        const pipeline = this._adamwPipeline;
-        const params = this.model.parameters();
-        const { learningRate: lr, weightDecay: wd, beta1, beta2, eps, beta1_t, beta2_t } = hp;
-        for (let pi = 0; pi < params.length; pi++) {
-            const p = params[pi];
-            const mom = this._moments[pi];
-            const paramBuf = createStorageBuffer(device, p.data, true);
-            const gradBuf = createStorageBuffer(device, grads[pi].data, false);
-            const mBuf = createStorageBuffer(device, mom.m, true);
-            const vBuf = createStorageBuffer(device, mom.v, true);
-            const uni = createUniformBuffer(device, packAdamParams(p.numel, lr, beta1, beta2, eps, wd, beta1_t, beta2_t));
-            const bg = createBindGroup(device, pipeline, [uni, paramBuf, gradBuf, mBuf, vBuf]);
-            dispatchKernel(device, pipeline, bg, [cdiv(p.numel, 256), 1, 1]);
-            p.data.set((await readBuffer(device, paramBuf, p.numel * 4)).subarray(0, p.numel));
-            mom.m.set((await readBuffer(device, mBuf, p.numel * 4)).subarray(0, p.numel));
-            mom.v.set((await readBuffer(device, vBuf, p.numel * 4)).subarray(0, p.numel));
-            paramBuf.destroy();
-            gradBuf.destroy();
-            mBuf.destroy();
-            vBuf.destroy();
-            uni.destroy();
-        }
+        return this._trainer.evaluate(samples);
     }
 }
 //# sourceMappingURL=limbic_trainer.js.map

package/dist/limbic/limbic_trainer.js.map

@@ -1 +1 @@
-{"version":3,"file":"limbic_trainer.js","sourceRoot":"","sources":["../../src/limbic/limbic_trainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EACL,mBAAmB,EACnB,mBAAmB,EACnB,qBAAqB,EACrB,eAAe,EACf,cAAc,EACd,UAAU,EACV,IAAI,GACL,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAE,kBAAkB,EAAE,MAAM,6BAA6B,CAAC;AAmCjE,SAAS,cAAc,CACrB,WAAmB,EACnB,EAAU,EACV,KAAa,EACb,KAAa,EACb,GAAW,EACX,WAAmB,EACnB,OAAe,EACf,OAAe;IAEf,MAAM,GAAG,GAAG,IAAI,WAAW,CAAC,EAAE,CAAC,CAAC;IAChC,IAAI,WAAW,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC;IAC9C,IAAI,YAAY,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,WAAW,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;IACxF,OAAO,GAAG,CAAC;AACb,CAAC;AAED,MAAM,OAAO,aAAa;IACf,KAAK,CAAc;IACnB,MAAM,CAAmB;IAC1B,QAAQ,GAAwB,IAAI,CAAC;IACrC,KAAK,GAAG,CAAC,CAAC;IACD,cAAc,CAA4B;IAE3D,YAAY,KAAkB,EAAE,SAA2B,IAAI;QAC7D,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,cAAc,GAAG,MAAM;YAC1B,CAAC,CAAC,qBAAqB,CAAC,MAAM,EAAE,kBAAkB,EAAE,cAAc,CAAC;YACnE,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;IAED,kDAAkD;IAClD,IAAI,WAAW;QACb,OAAO,IAAI,CAAC,MAAM,IAAI,IAAI,IAAI,IAAI,CAAC,cAAc,IAAI,IAAI,CAAC;IAC5D,CAAC;IAEO,YAAY;QAClB,IAAI,IAAI,CAAC,QAAQ;YAAE,OAAO;QAC1B,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YAClD,CAAC,EAAE,IAAI,YAAY,CAAC,CAAC,CAAC,KAAK,CAAC;YAC5B,CAAC,EAAE,IAAI,YAAY,CAAC,CAAC,CAAC,KAAK,CAAC;SAC7B,CAAC,CAAC,CAAC;IACN,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,KAAK,CAAC,OAAuB,EAAE,OAA2B,EAAE;QAChE,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,iCAAiC,CAAC,CAAC;QAC7E,MAAM,EACJ,YAAY,GAAG,IAAI,EACnB,MAAM,GAAG,EAAE,EACX,WAAW,GAAG,GAAG,EACjB,KAAK,GAAG,GAAG,EACX,KAAK,GAAG,KAAK,EACb,GAAG,GAAG,IAAI,EACV,WAAW,GAAG,GAAG,EACjB,UAAU,GAAG,IAAI,GAClB,GAAG,IAAI,CAAC;QAET,IAAI,CAAC,YAAY,EAAE,CAAC;QACpB,MAAM,MAAM,GAAa,EAAE,CAAC;QAE5B,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;YAC5C,IAAI,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC;YACtB,IAAI,SAAS,GAAG,CAAC,CAAC;YAElB,yEAAyE;YACzE,yEAAyE;YACzE,2EAA2E;YAC3E,qDAAqD;YACrD,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;gBACxB,MAAM,EAAE,IAAI,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CACtC,CAAC,CAAC,KAAK,EACP,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,EACvB,CAAC,CAAC,KAAK,EACP,CAAC,CAAC,WAAW,EACb,CAAC,CAAC,MAAM,CACT,CAAC;gBACF,SAAS,IAAI,IAAI,CAAC;YACpB,CAAC;YAED,oCAAoC;YACpC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,EAAE,CAAC;YACrC,MAAM,IAAI,GAAG,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;YAChC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;gBACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE;oBAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAE,IAAI,IAAI,CAAC;YAC7D,CAAC;YAED,MAAM,QAAQ,GAAG,IAAI,CAAC,cAAc,CAAC,KAAK,EAAE,WAAW,CAAC,CAAC;YAEzD,IAAI,CAAC,KAAK,EAAE,CAAC;YACb,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;YAC5C,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;YAC5C,MAAM,EAAE,GAAG,EAAE,YAAY,EAAE,WAAW,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;YAC9E,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;gBACrB,MAAM,IAAI,CAAC,aAAa,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;YACtC,CAAC;iBAAM,CAAC;gBACN,IAAI,CAAC,aAAa,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;YAChC,CAAC;YAED,MAAM,GAAG,GAAG,SAAS,GAAG,OAAO,CAAC,MAAM,CAAC;YACvC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YACjB,IAAI,UAAU;gBAAE,UAAU,CAAC,KAAK,GAAG,CAAC,EAAE,GAAG,EAAE,QAAQ,CAAC,CAAC;QACvD,CAAC;QACD,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,+EAA+E;IAC/E,QAAQ,CAAC,OAAuB;QAC9B,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,CAAC,CAAC;QACnC,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,MAAM,EAAE,YAAY,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;QAC3C,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;YACxB,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC;YACxE,IAAI,IAAI,GAAG,CAAC,CAAC;YACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACxC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;gBAChD,IAAI,IAAI,GAAG,GAAG,CAAC,GAAG,CAAC,CAAC;YACtB,CAAC;YACD,MAAM,EAAE,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,MAAM,CAAC;YAC/B,IAAI,IAAI,GAAG,GAAG,YAAY,GAAG,EAAE,GAAG,EAAE,CAAC;YACrC,KAAK,IAAI,IAAI,CAAC;QAChB,CAAC;QACD,OAAO,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC;IAChC,CAAC;IAED,mFAAmF;IAC3E,cAAc,CAAC,KAAoB,EAAE,OAAe;QAC1D,IAAI,MAAM,GAAG,CAAC,CAAC;QACf,KAAK,MAAM,CAAC,IAAI,KAAK;YAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE;gBAAE,MAAM,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAE,CAAC;QACjG,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAC/B,IAAI,IAAI,GAAG,OAAO,IAAI,IAAI,GAAG,CAAC,EAAE,CAAC;YAC/B,MAAM,KAAK,GAAG,OAAO,GAAG,IAAI,CAAC;YAC7B,KAAK,MAAM,CAAC,IAAI,KAAK;gBAAE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE;oBAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAE,IAAI,KAAK,CAAC;QACrF,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;IAEO,aAAa,CACnB,KAAoB,EACpB,EAA8H;QAE9H,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC;QACvC,MAAM,EAAE,YAAY,EAAE,EAAE,EAAE,WAAW,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,CAAC;QACtF,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,EAAE,EAAE,EAAE,CAAC;YAC1C,MAAM,CAAC,GAAG,MAAM,CAAC,EAAE,CAAE,CAAC,IAAI,CAAC;YAC3B,MAAM,CAAC,GAAG,KAAK,CAAC,EAAE,CAAE,CAAC,IAAI,CAAC;YAC1B,MAAM,GAAG,GAAG,IAAI,CAAC,QAAS,CAAC,EAAE,CAAE,CAAC;YAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC;gBACjB,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,CAAC;gBAChD,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,EAAE,GAAG,EAAE,CAAC;gBACrD,MAAM,IAAI,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC;gBACvC,MAAM,IAAI,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC;gBACvC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,GAAG,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,CAAC,CAAC;YACvE,CAAC;QACH,CAAC;IACH,CAAC;IAED,mFAAmF;IAC3E,KAAK,CAAC,aAAa,CACzB,KAAoB,EACpB,EAA8H;QAE9H,MAAM,MAAM,GAAG,IAAI,CAAC,MAAO,CAAC;QAC5B,MAAM,QAAQ,GAAG,IAAI,CAAC,cAAe,CAAC;QACtC,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC;QACvC,MAAM,EAAE,YAAY,EAAE,EAAE,EAAE,WAAW,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,CAAC;QAEtF,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,EAAE,EAAE,EAAE,CAAC;YAC1C,MAAM,CAAC,GAAG,MAAM,CAAC,EAAE,CAAE,CAAC;YACtB,MAAM,GAAG,GAAG,IAAI,CAAC,QAAS,CAAC,EAAE,CAAE,CAAC;YAChC,MAAM,QAAQ,GAAG,mBAAmB,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;YAC3D,MAAM,OAAO,GAAG,mBAAmB,CAAC,MAAM,EAAE,KAAK,CAAC,EAAE,CAAE,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;YACpE,MAAM,IAAI,GAAG,mBAAmB,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;YACtD,MAAM,IAAI,GAAG,mBAAmB,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;YACtD,MAAM,GAAG,GAAG,mBAAmB,CAC7B,MAAM,EACN,cAAc,CAAC,CAAC,CAAC,KAAK,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE,EAAE,OAAO,EAAE,OAAO,CAAC,CACrE,CAAC;YACF,MAAM,EAAE,GAAG,eAAe,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;YACnF,cAAc,CAAC,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,EAAE,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;YAEjE,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,MAAM,UAAU,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;YACnF,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,UAAU,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;YAC9E,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,UAAU,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;YAE9E,QAAQ,CAAC,OAAO,EAAE,CAAC;YACnB,OAAO,CAAC,OAAO,EAAE,CAAC;YAClB,IAAI,CAAC,OAAO,EAAE,CAAC;YACf,IAAI,CAAC,OAAO,EAAE,CAAC;YACf,GAAG,CAAC,OAAO,EAAE,CAAC;QAChB,CAAC;IACH,CAAC;CACF"}
+{"version":3,"file":"limbic_trainer.js","sourceRoot":"","sources":["../../src/limbic/limbic_trainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,WAAW,EAA6C,MAAM,uBAAuB,CAAC;AAkB/F;;;;;GAKG;AACH,MAAM,UAAU,UAAU,CAAC,KAAkB;IAC3C,OAAO;QACL,UAAU,EAAE,GAAkB,EAAE,CAAC,KAAK,CAAC,UAAU,EAAE;QACnD,SAAS,EAAE,GAAkB,EAAE,CAAC,KAAK,CAAC,SAAS,EAAE;QACjD,QAAQ,EAAE,GAAG,EAAE,CAAC,KAAK,CAAC,QAAQ,EAAE;QAChC,UAAU,EAAE,CAAC,CAAC,EAAE,EAAE,CAChB,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,KAAK,EAAE,KAAK,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI;QACxF,IAAI,EAAE,CAAC,CAAC,EAAE,EAAE;YACV,MAAM,EAAE,YAAY,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC;YACtC,MAAM,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,EAAE,KAAK,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC;YAC9D,IAAI,IAAI,GAAG,CAAC,CAAC;YACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACxC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAE,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;gBAChD,IAAI,IAAI,GAAG,GAAG,CAAC,GAAG,CAAC,CAAC;YACtB,CAAC;YACD,MAAM,EAAE,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,MAAM,CAAC;YAC/B,IAAI,IAAI,GAAG,GAAG,YAAY,GAAG,EAAE,GAAG,EAAE,CAAC;YACrC,OAAO,IAAI,CAAC;QACd,CAAC;KACF,CAAC;AACJ,CAAC;AAED,gFAAgF;AAChF,MAAM,OAAO,aAAa;IACf,KAAK,CAAc;IACnB,MAAM,CAAmB;IACjB,QAAQ,CAA4B;IAErD,YAAY,KAAkB,EAAE,SAA2B,IAAI;QAC7D,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,QAAQ,GAAG,IAAI,WAAW,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,MAAM,EAAE,eAAe,CAAC,CAAC;IAC9E,CAAC;IAED,kDAAkD;IAClD,IAAI,WAAW;QACb,OAAO,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC;IACnC,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,OAAuB,EAAE,OAA2B,EAAE;QAC1D,OAAO,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC;IAC5C,CAAC;IAED,+EAA+E;IAC/E,QAAQ,CAAC,OAAuB;QAC9B,OAAO,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;IACzC,CAAC;CACF"}

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

@@ -0,0 +1,32 @@
+/**
+ * ml/classical/ewma.ts — exponentially-weighted moving average + variance, for
+ * drift anomalies in an ordered series (a run's cost creeping up over the week).
+ *
+ * Per feature, over the history IN ORDER: diff = x − μ; μ += α·diff;
+ * σ² = (1 − α)(σ² + α·diff²). A new point scores z = |x − μ| / σ against the
+ * state BEFORE it is folded in. Prediction: `scores[0] = max_i z_i / threshold`.
+ */
+import type { MlPrediction, MlPredictor, MlSample } from "../types.js";
+export interface EwmaOptions {
+    /** Smoothing factor in (0, 1]. Default 0.3. */
+    alpha?: number;
+    /** z at which a feature counts as anomalous. Default 3. */
+    threshold?: number;
+    /** Smallest σ per feature, as a fraction of |μ| (plus 1e-9). Default 0.01. */
+    minScaleFraction?: number;
+}
+export declare class EwmaModel implements MlPredictor {
+    readonly mean: Float64Array;
+    readonly variance: Float64Array;
+    readonly alpha: number;
+    readonly threshold: number;
+    readonly minScaleFraction: number;
+    constructor(mean: Float64Array, variance: Float64Array, alpha: number, threshold: number, minScaleFraction: number);
+    zScores(x: ArrayLike<number>): number[];
+    /** Fold one observation into the running state. */
+    update(x: ArrayLike<number>): void;
+    predict(x: Float32Array): MlPrediction;
+}
+/** Fit over `samples` in the given (chronological) order. */
+export declare function fitEwma(samples: readonly MlSample[], opts?: EwmaOptions): EwmaModel;
+//# sourceMappingURL=ewma.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"ewma.d.ts","sourceRoot":"","sources":["../../../src/ml/classical/ewma.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAIvE,MAAM,WAAW,WAAW;IAC1B,+CAA+C;IAC/C,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,2DAA2D;IAC3D,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,8EAA8E;IAC9E,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED,qBAAa,SAAU,YAAW,WAAW;IAEzC,QAAQ,CAAC,IAAI,EAAE,YAAY;IAC3B,QAAQ,CAAC,QAAQ,EAAE,YAAY;IAC/B,QAAQ,CAAC,KAAK,EAAE,MAAM;IACtB,QAAQ,CAAC,SAAS,EAAE,MAAM;IAC1B,QAAQ,CAAC,gBAAgB,EAAE,MAAM;gBAJxB,IAAI,EAAE,YAAY,EAClB,QAAQ,EAAE,YAAY,EACtB,KAAK,EAAE,MAAM,EACb,SAAS,EAAE,MAAM,EACjB,gBAAgB,EAAE,MAAM;IAKnC,OAAO,CAAC,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,GAAG,MAAM,EAAE;IAOvC,mDAAmD;IACnD,MAAM,CAAC,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,GAAG,IAAI;IASlC,OAAO,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY;CAIvC;AAED,6DAA6D;AAC7D,wBAAgB,OAAO,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE,EAAE,IAAI,GAAE,WAAgB,GAAG,SAAS,CAWvF"}

package/dist/ml/classical/ewma.js

@@ -0,0 +1,46 @@
+import { featureDim } from "./math.js";
+import { anomalyConfidence } from "./robust_z.js";
+export class EwmaModel {
+    mean;
+    variance;
+    alpha;
+    threshold;
+    minScaleFraction;
+    constructor(mean, variance, alpha, threshold, minScaleFraction) {
+        this.mean = mean;
+        this.variance = variance;
+        this.alpha = alpha;
+        this.threshold = threshold;
+        this.minScaleFraction = minScaleFraction;
+        if (!(alpha > 0 && alpha <= 1))
+            throw new Error(`EwmaModel: alpha must be in (0, 1] (got ${alpha})`);
+    }
+    zScores(x) {
+        return Array.from(this.mean, (m, i) => {
+            const sd = Math.max(Math.sqrt(this.variance[i]), this.minScaleFraction * Math.abs(m) + 1e-9);
+            return Math.abs((x[i] ?? 0) - m) / sd;
+        });
+    }
+    /** Fold one observation into the running state. */
+    update(x) {
+        for (let i = 0; i < this.mean.length; i++) {
+            const diff = (x[i] ?? 0) - this.mean[i];
+            const incr = this.alpha * diff;
+            this.mean[i] = this.mean[i] + incr;
+            this.variance[i] = (1 - this.alpha) * (this.variance[i] + diff * incr);
+        }
+    }
+    predict(x) {
+        const score = Math.max(0, ...this.zScores(x)) / this.threshold;
+        return { scores: [score], confidence: anomalyConfidence(score) };
+    }
+}
+/** Fit over `samples` in the given (chronological) order. */
+export function fitEwma(samples, opts = {}) {
+    const d = featureDim(samples.map((s) => s.x));
+    const model = new EwmaModel(Float64Array.from(samples[0].x), new Float64Array(d), opts.alpha ?? 0.3, opts.threshold ?? 3, opts.minScaleFraction ?? 0.01);
+    for (let i = 1; i < samples.length; i++)
+        model.update(samples[i].x);
+    return model;
+}
+//# sourceMappingURL=ewma.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"ewma.js","sourceRoot":"","sources":["../../../src/ml/classical/ewma.ts"],"names":[],"mappings":"AASA,OAAO,EAAE,UAAU,EAAE,MAAM,WAAW,CAAC;AACvC,OAAO,EAAE,iBAAiB,EAAE,MAAM,eAAe,CAAC;AAWlD,MAAM,OAAO,SAAS;IAET;IACA;IACA;IACA;IACA;IALX,YACW,IAAkB,EAClB,QAAsB,EACtB,KAAa,EACb,SAAiB,EACjB,gBAAwB;QAJxB,SAAI,GAAJ,IAAI,CAAc;QAClB,aAAQ,GAAR,QAAQ,CAAc;QACtB,UAAK,GAAL,KAAK,CAAQ;QACb,cAAS,GAAT,SAAS,CAAQ;QACjB,qBAAgB,GAAhB,gBAAgB,CAAQ;QAEjC,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,2CAA2C,KAAK,GAAG,CAAC,CAAC;IACvG,CAAC;IAED,OAAO,CAAC,CAAoB;QAC1B,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;YACpC,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,EAAE,IAAI,CAAC,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC;YAC9F,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC;QACxC,CAAC,CAAC,CAAC;IACL,CAAC;IAED,mDAAmD;IACnD,MAAM,CAAC,CAAoB;QACzB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC1C,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAE,CAAC;YACzC,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC;YAC/B,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAE,GAAG,IAAI,CAAC;YACpC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAE,GAAG,IAAI,GAAG,IAAI,CAAC,CAAC;QAC1E,CAAC;IACH,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,6DAA6D;AAC7D,MAAM,UAAU,OAAO,CAAC,OAA4B,EAAE,OAAoB,EAAE;IAC1E,MAAM,CAAC,GAAG,UAAU,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC9C,MAAM,KAAK,GAAG,IAAI,SAAS,CACzB,YAAY,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,EAChC,IAAI,YAAY,CAAC,CAAC,CAAC,EACnB,IAAI,CAAC,KAAK,IAAI,GAAG,EACjB,IAAI,CAAC,SAAS,IAAI,CAAC,EACnB,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAC9B,CAAC;IACF,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,KAAK,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC;IACrE,OAAO,KAAK,CAAC;AACf,CAAC"}

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

@@ -0,0 +1,42 @@
+/**
+ * ml/classical/gbm.ts — a small gradient-boosted tree ensemble.
+ *
+ * Depth-limited regression trees fitted to the loss gradient, one after another,
+ * shrunk by a learning rate. Objectives: `binary` (log-loss, Newton leaf values)
+ * and `regression` (squared error). Exhaustive split search over midpoints —
+ * fine for the catalog's small numeric feature sets, not for wide embeddings.
+ */
+import type { MlPrediction, MlPredictor, MlSample } from "../types.js";
+import { type Vec } from "./math.js";
+export type GbmObjective = "binary" | "regression";
+export interface GbmOptions {
+    objective?: GbmObjective;
+    trees?: number;
+    maxDepth?: number;
+    learningRate?: number;
+    /** Minimum summed sample weight in a leaf. Default 1. */
+    minLeafWeight?: number;
+    /** L2 on leaf values (Newton denominator). Default 1. */
+    lambda?: number;
+}
+type RTree = {
+    value: number;
+} | {
+    feature: number;
+    split: number;
+    left: RTree;
+    right: RTree;
+};
+export declare class GbmModel implements MlPredictor {
+    readonly objective: GbmObjective;
+    readonly base: number;
+    readonly trees: readonly RTree[];
+    readonly learningRate: number;
+    constructor(objective: GbmObjective, base: number, trees: readonly RTree[], learningRate: number);
+    /** The raw additive score (log-odds for `binary`). */
+    margin(x: Vec): number;
+    predict(x: Float32Array): MlPrediction;
+}
+export declare function fitGbm(samples: readonly MlSample[], opts?: GbmOptions): GbmModel;
+export {};
+//# sourceMappingURL=gbm.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"gbm.d.ts","sourceRoot":"","sources":["../../../src/ml/classical/gbm.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AACvE,OAAO,EAAuB,KAAK,GAAG,EAAE,MAAM,WAAW,CAAC;AAE1D,MAAM,MAAM,YAAY,GAAG,QAAQ,GAAG,YAAY,CAAC;AAEnD,MAAM,WAAW,UAAU;IACzB,SAAS,CAAC,EAAE,YAAY,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,yDAAyD;IACzD,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,yDAAyD;IACzD,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,KAAK,KAAK,GAAG;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,KAAK,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC;AAkE/F,qBAAa,QAAS,YAAW,WAAW;IAExC,QAAQ,CAAC,SAAS,EAAE,YAAY;IAChC,QAAQ,CAAC,IAAI,EAAE,MAAM;IACrB,QAAQ,CAAC,KAAK,EAAE,SAAS,KAAK,EAAE;IAChC,QAAQ,CAAC,YAAY,EAAE,MAAM;gBAHpB,SAAS,EAAE,YAAY,EACvB,IAAI,EAAE,MAAM,EACZ,KAAK,EAAE,SAAS,KAAK,EAAE,EACvB,YAAY,EAAE,MAAM;IAG/B,sDAAsD;IACtD,MAAM,CAAC,CAAC,EAAE,GAAG,GAAG,MAAM;IAMtB,OAAO,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY;CAMvC;AAED,wBAAgB,MAAM,CAAC,OAAO,EAAE,SAAS,QAAQ,EAAE,EAAE,IAAI,GAAE,UAAe,GAAG,QAAQ,CA6BpF"}