import type { TaskClass } from "./types.ts";
/**
 * A point estimate plus its 95% Wilson-score confidence interval.
 *
 * - `mean`: best-estimate success probability in [0, 1].
 * - `lo95` / `hi95`: 95% Wilson-score interval bounds in [0, 1].
 * - `n`: number of evaluation trials backing the estimate. When the
 *   measurement file is missing this equals `PRIOR_N` (see below) so
 *   callers can detect the fallback case.
 */
export interface QualityWithCI {
    mean: number;
    lo95: number;
    hi95: number;
    n: number;
}
/**
 * Synthetic trial count assigned to a fallback-prior cell so that the CI
 * is well-defined. Chosen at 10 trials, which yields a wide CI (~+/-0.3
 * for p=0.5) — that's correct behaviour: the prior should be treated as
 * weak evidence, and a caller routing on `lo95` will be appropriately
 * conservative until real measurements land.
 */
export declare const PRIOR_N = 10;
/**
 * Wilson score 95% confidence interval for a binomial proportion.
 *
 * Reference: Wilson, E. B. (1927). "Probable inference, the law of
 * succession, and statistical inference." JASA 22(158): 209-212.
 *
 * Why Wilson over Wald: Wald (the textbook normal approximation, p̂ ±
 * z*sqrt(p̂(1-p̂)/n)) collapses to the empty interval when p̂ ∈ {0, 1}
 * and underestimates uncertainty for small n. Wilson is well-defined at
 * the boundary, has better coverage at small n, and is the recommended
 * default for binomial CIs in modern stats texts (e.g. Agresti & Coull,
 * 1998 — "Approximate is better than 'exact' for interval estimation of
 * binomial proportions").
 *
 * Inputs:
 *   - `successes`: integer in [0, trials].
 *   - `trials`: positive integer.
 *
 * Returns the (lo, hi) tuple clamped to [0, 1]. Pure function.
 */
export declare function wilsonScore95(successes: number, trials: number): {
    lo: number;
    hi: number;
};
/**
 * Test-only hook: clears the memoized quality table so the next lookup
 * re-reads disk. Mirrors `__resetCalibrationCacheForTest` in `cost.ts`.
 * Not part of the public API surface but intentionally exported with the
 * `__`-prefix convention so tests can reach for it explicitly.
 */
export declare const __resetQualityCacheForTest: () => void;
/**
 * Return the expected quality for `(taskClass, modelId)` as a point estimate
 * in [0, 1].
 *
 * Backward-compatible signature: matches `quality_prior.predictQuality` so
 * the router doesn't need to change. The mean comes from
 * `predictQualityWithCI` — callers wanting uncertainty should use that
 * function directly.
 *
 * Unknown models receive `DEFAULT_QUALITY = 0.5`.
 */
export declare function predictQuality(taskClass: TaskClass, modelId: string): number;
/**
 * Return the expected quality for `(taskClass, modelId)` with a 95%
 * Wilson-score confidence interval.
 *
 * - `mean` is `successes / trials`.
 * - `lo95` / `hi95` is the Wilson score interval at the 95% level.
 * - `n` is the trial count (real for measured cells, `PRIOR_N` for the
 *   seeded prior fallback).
 *
 * Unknown (taskClass, modelId) pairs return a uniform-prior estimate:
 *   `{ mean: 0.5, lo95, hi95, n: PRIOR_N }`
 * where the CI is the Wilson 95% interval for 5 successes in 10 trials —
 * intentionally wide to reflect that no data backs the estimate.
 */
export declare function predictQualityWithCI(taskClass: TaskClass, modelId: string): QualityWithCI;
/**
 * Introspection helper. Returns whether the predictor is currently serving
 * measurements or the fallback prior, and the metadata of the loaded file.
 * Useful for the CLI's `route --debug` output and for the paper's
 * reproducibility appendix.
 */
export declare function getQualitySourceInfo(): {
    source: "measured" | "prior";
    loadedFrom?: string;
    generatedAt?: string;
};
//# sourceMappingURL=quality_predictor.d.ts.map