/**
 * The failure that actually matters, and the four ways of getting it wrong.
 *
 * CI has been able to fail a build for tokens since 1.4 and for dollars since
 * 1.21. A prompt edit that quietly made the product worse has never been
 * gateable — which means every cost saving this tool has ever recommended went
 * into a repository with its most important consequence unmeasured.
 *
 * ## This is a before-and-after, not an experiment
 *
 * Worth saying in the first paragraph, because the arithmetic is the same as
 * `experiment`'s and the epistemics are not. An experiment splits traffic
 * randomly, so the two arms differ only in the thing under test. A
 * before-and-after splits it by **time**, and everything else that changed at
 * the same time is in the difference too.
 *
 * So this module spends most of its code looking for reasons *not* to blame the
 * prompt, and it reports `cannot-tell` far more readily than a randomised
 * comparison would. That is not timidity. A gate that blames the prompt because
 * the prompt is the thing it can see will be switched off within a month, and a
 * switched-off gate catches nothing at all.
 *
 * ## The three confounders it can actually see
 *
 * **The model moved.** If the mix of models shifted across the boundary, the
 * prompt is not the only variable, and the drop may be entirely somebody else's
 * migration.
 *
 * **The volume moved.** A workload whose traffic tripled is usually a workload
 * whose *population* changed — a new surface, a new customer, a marketing
 * campaign — and the questions being asked are not the questions from before.
 *
 * **The coverage moved.** The subtle one, and the one nobody thinks of: if the
 * share of calls recording an outcome changed, the two rates describe different
 * populations. A team that started instrumenting its hard cases will see its
 * measured rate fall without anything getting worse.
 *
 * Any of them present, and the verdict is `cannot-tell` with the confounder
 * **named**. Not a hedge attached to a blame — a refusal to blame.
 *
 * ## What it cannot see, and says so
 *
 * Everything else deployed that day. This module has one label's numbers and no
 * knowledge of the world, and it never pretends otherwise: a `dropped` verdict
 * is a statement that the rate fell and the three things it can check did not
 * move, which is a smaller claim than "the prompt did it" and is the largest
 * claim the evidence supports.
 */
import type { ArmResult, ExperimentArm } from './experiment.js';
import type { OutcomeVocabulary } from './outcome.js';
/**
 * Outcomes each side of the change needs before a verdict is offered.
 *
 * A hundred rather than the ten used for a rate elsewhere, because this one
 * fails builds. The cost of a wrong `dropped` is somebody reverting a good
 * change and losing the saving; the cost of a wrong `cannot-tell` is waiting a
 * day. Those are not symmetric and the threshold is not either.
 */
export declare const MIN_OUTCOMES_EACH_SIDE = 100;
/** How far the model mix may move before the comparison stops being about the prompt. */
export declare const MAX_MODEL_MIX_DRIFT = 0.1;
/**
 * How far the call volume may move, as a ratio.
 *
 * Half again, either way. Ordinary week-to-week traffic moves less than that;
 * a workload that doubled is usually a workload whose population changed.
 */
export declare const MAX_VOLUME_RATIO = 1.5;
/** How far outcome coverage may move before the two rates describe different populations. */
export declare const MAX_COVERAGE_DRIFT = 0.1;
export type Confounder = {
    kind: 'model-mix-moved';
    drift: number;
    model: string;
} | {
    kind: 'volume-moved';
    beforeCalls: number;
    afterCalls: number;
    ratio: number;
} | {
    kind: 'coverage-moved';
    before: number;
    after: number;
};
export type GateVerdict = 'held' | 'dropped' | 'cannot-tell';
export type GateUnknown = 'too-few-before' | 'too-few-after' | 'confounded' | 'no-vocabulary' | 'not-separable';
/** One side of the boundary, as the caller measured it. */
export interface GateSide {
    arm: ExperimentArm;
    /** Every call in the window, outcome or not — the coverage denominator. */
    calls: number;
    /** Spend per model, for the mix comparison. */
    usdByModel: Array<{
        model: string;
        usd: number;
    }>;
}
export interface QualityGate {
    verdict: GateVerdict;
    /** Why, when the verdict is `cannot-tell`. A refusal never arrives bare. */
    unknown: GateUnknown | null;
    before: ArmResult;
    after: ArmResult;
    /** The 95% interval on (after − before), or null. */
    difference: {
        point: number;
        low: number;
        high: number;
    } | null;
    /**
     * Everything found that means the prompt is not the only variable.
     *
     * Non-empty forces `cannot-tell`. Reported even on a `held` verdict, because
     * a rate that held while the model changed underneath is not evidence that
     * the prompt is fine either.
     */
    confounders: Confounder[];
    /**
     * The cost half — measured, and kept beside the quality half rather than
     * merged with it. The sentence teams argue about needs both.
     */
    cost: {
        beforeUsdPerCall: number;
        afterUsdPerCall: number;
        deltaUsdPerCall: number;
    } | null;
    /** Outcomes behind the comparison, so the claim carries its own sample size. */
    outcomes: {
        before: number;
        after: number;
    };
}
export declare function qualityGate(before: GateSide, after: GateSide, vocabulary: OutcomeVocabulary | null): QualityGate;
//# sourceMappingURL=quality-gate.d.ts.map