/**
 * Sense 1: Temporal Fingerprint — PRIMARY (88.5% standalone accuracy, 5x weight)
 *
 * The model's inference rhythm. Token-by-token timing, inter-token intervals,
 * burst patterns, time-to-first-token. This is the model's heartbeat.
 *
 * Every model architecture produces a distinctive timing signature determined
 * by its weights, attention mechanism, and hardware. You cannot produce Claude's
 * rhythm without running Claude's inference. This is the voice.
 *
 * Beyond aggregate statistics, the temporal sense also measures timing
 * *conditioned on context* — the Conditional Timing Surface. How long the model
 * takes to generate token N depends on the attention computation over tokens
 * 1 through N-1. Different weights -> different attention patterns -> different
 * per-token timing conditioned on the specific context.
 *
 * This creates a high-dimensional conditional timing surface that a distilled
 * model cannot replicate without equivalent computation on equivalent hardware.
 */
import type { TokenStreamCapture } from "../stream-capture.js";
export interface TemporalSignals {
    /** ms from request to first token. */
    timeToFirstToken: number;
    /** Average inter-token interval in ms. */
    meanInterval: number;
    /** Median inter-token interval in ms. */
    medianInterval: number;
    /** Standard deviation of inter-token intervals. */
    stdInterval: number;
    /** Burstiness = variance / mean of intervals. Measures batchy vs smooth delivery. */
    burstiness: number;
    /** Total generation time in ms. */
    totalStreamingDuration: number;
    /** Number of tokens generated. */
    tokenCount: number;
    /** Mean interval in the first quartile of tokens. */
    earlyMeanInterval: number;
    /** Mean interval in the middle half of tokens. */
    midMeanInterval: number;
    /** Mean interval in the last quartile of tokens. */
    lateMeanInterval: number;
    /** Timing acceleration: late mean - early mean. Positive = slowing down. */
    acceleration: number;
    /** Ratio of avg interval at sentence boundaries vs mid-sentence. */
    sentenceBoundaryPauseRatio: number;
    /** Number of detected bursts. */
    burstCount: number;
    /** Fraction of tokens that are part of bursts. */
    burstFraction: number;
    /** Average number of tokens per burst. */
    meanBurstLength: number;
    /** Mean interval between consecutive bursts (ms). */
    interBurstInterval: number;
    /** Number of inference chunks (network-level). */
    chunkCount: number;
    /** Average tokens per chunk. */
    meanChunkSize: number;
    /** Shannon entropy of binned interval distribution. */
    timingEntropy: number;
    /** 90th percentile inter-token interval. */
    p90Interval: number;
    /** 10th percentile inter-token interval. */
    p10Interval: number;
    /** Skewness of the interval distribution. */
    intervalSkewness: number;
}
export declare const TEMPORAL_FEATURE_NAMES: readonly string[];
/**
 * Extract temporal signals from a token stream capture.
 * This is the PRIMARY sense — 88.5% standalone accuracy.
 */
export declare function extractTemporalSignals(capture: TokenStreamCapture): TemporalSignals;
/**
 * Convert temporal signals to a feature vector in the order defined by TEMPORAL_FEATURE_NAMES.
 */
export declare function temporalToFeatureVector(signals: TemporalSignals): number[];
//# sourceMappingURL=temporal.d.ts.map