/**
 * bench/metrics.ts — pure scoring primitives over per-position logits.
 *
 * These are the building blocks the harness composes. They are exact, allocation
 * -light, and numerically stable (cross-entropy reuses the engine's stable
 * log-sum-exp). Nothing here trains or mutates the model — benchmarking is a
 * read-only forward pass over held-out data.
 */
/** Natural log of 2 — used to convert nats → bits. */
export declare const LN2: number;
/**
 * Indices of the `k` largest logits, descending. O(V·k) — fine for the small
 * top-k used in accuracy scoring; avoids a full sort of the vocab per position.
 */
export declare function topKIndices(logits: Float32Array, k: number): number[];
/** Index of the single largest logit (argmax). */
export declare function argmax(logits: Float32Array): number;
/** Perplexity from a mean cross-entropy (nats). */
export declare function perplexity(meanCrossEntropy: number): number;
/** Bits-per-token from a mean cross-entropy (nats). */
export declare function bitsPerToken(meanCrossEntropy: number): number;
/** Accumulator over predicted positions — cross-entropy + top-1/top-k hits. */
export interface ScoreAccumulator {
    tokens: number;
    ceSum: number;
    top1Hits: number;
    topKHits: number;
}
export declare function newAccumulator(): ScoreAccumulator;
/**
 * Score one (logits, targets) pair into the accumulator. `logits[t]` predicts
 * `targets[t]`; positions with an out-of-range target are skipped. The caller
 * aligns logits/targets (next-token: logits[0..T-2] predict tokens[1..T-1]).
 */
export declare function scoreInto(acc: ScoreAccumulator, logits: Float32Array[], targets: number[], k: number): void;
//# sourceMappingURL=metrics.d.ts.map