/**
 * EffectivenessTracker — the closed adaptive loop (adaptive-agent design R4, leapfrog #9).
 *
 * Recall (R3) injects a `<memory_context>` page; this tracks whether the agent actually USED it, so the
 * recall gate can adapt — recall more when it's paying off, back off when it isn't. "Used" = the
 * fraction of the recall page's DISTINCTIVE tokens (those not already in the user's query) that reappear
 * in the assistant's response. We isolate distinctive tokens so we measure recall's own contribution,
 * not the baseline overlap every response shares with the query.
 *
 * The score is an exponential moving average ("useful lately") in [0,1], starting at a neutral prior so
 * recall is given a fair chance before the loop adapts.
 */
export declare class EffectivenessTracker {
    private ema;
    private samples;
    /**
     * Record the outcome of a turn that received a recall page: how much of the recall's distinctive
     * content the assistant's response actually drew on.
     */
    recordRecallOutcome(recallText: string, queryText: string, responseText: string): void;
    /** Rolling "useful lately" score in [0,1]. Neutral until enough samples accumulate. */
    usefulLately(): number;
    /** Number of recorded recall outcomes. */
    get sampleCount(): number;
}
/**
 * Fraction of the recall page's distinctive tokens (present in recall but NOT in the query) that appear
 * in the response. 0 when recall added nothing the query didn't already carry.
 */
export declare function distinctiveRecallUsage(recallText: string, queryText: string, responseText: string): number;
//# sourceMappingURL=effectiveness-tracker.d.ts.map