/**
 * COUNTING THE WITHHELD UNITS A REQUEST HAS BROUGHT BACK.
 *
 * Withholding a unit saves its residency for every remaining turn and costs
 * nothing unless it is wanted again, so the rate at which that happens is the
 * only thing the withholding case turns on. The denominator -- how many units
 * left -- has been counted at the spill sink all along. This is the numerator.
 *
 * It lives here rather than inside the proxy's closure so it can be tested
 * against a known answer. A counter that always reads zero passes every test
 * that only checks the proxy still runs, and zero is exactly the reading that
 * would make the arm look best.
 */
/** A digest of one withheld unit, as the spill sink records it. */
export declare const digestOf: (content: string) => string;
/**
 * How many withheld units this request brings back AS A RETRIEVAL.
 *
 * ONLY TOOL RESULTS COUNT, and the first version of this counted every block.
 * In a conversation the client resends the whole history every turn, so a unit
 * withheld at turn 5 arrives again at turns 6, 7 and 8 -- and counting those
 * made `reinstated` five times `spilled` over a real run, which is not a rate
 * at all. The client never reinstates anything here: it sends originals, the
 * proxy removes them on the way out, and the proxy re-decides every turn.
 *
 * What a retrieval actually looks like is the model calling the expand tool and
 * the content arriving as a NEW tool-result block. That is the only shape that
 * means somebody wanted the unit back, so it is the only shape counted.
 *
 * Counted once per unit however many times it appears, because the question is
 * whether it was wanted, not how often it was pasted. A body that does not
 * parse, or carries no messages, answers zero rather than throwing: a malformed
 * request is not evidence about the reference rate.
 */
export declare function reinstatedIn(body: string, withheld: ReadonlySet<string>): number;
/**
 * The reference rate: how often a withheld unit was wanted back.
 *
 * NULL WHEN NOTHING WAS WITHHELD, NEVER ZERO. Zero is the reading that makes
 * the withholding arm look best -- it says every unit was dropped for free --
 * and it is also what `0/0` produces when the arm never ran. Those two have to
 * be distinguishable, or a session with the arm switched off reports the most
 * flattering number available.
 *
 * The rate this returns is the whole subscription case: a withheld unit saves
 * its residency for every remaining turn and costs nothing unless it is wanted
 * again, so the arm's value is almost entirely a function of this one figure.
 * It has been estimated once, at 0.23, from textual recurrence on a corpus
 * belonging to the engine we are measuring against -- a method blind to a unit
 * the model read and reasoned about without quoting, so biased low, in the
 * direction that flatters us.
 */
export declare function referenceRate(counts: {
    readonly spilled?: number;
    readonly reinstated?: number;
}): number | null;
//# sourceMappingURL=reinstated.d.ts.map