/**
 * Pre-registered signals and the persistent regression suite.
 *
 * TWO DEFECTS IN THE SCOREBOARD THAT THIS FILE EXISTS TO FIX.
 *
 * 1. The scoreboard could only report NOVELTY, never REGRESSION. Every row was
 *    a point measurement of a different item, and nothing was ever re-measured.
 *    If item 1 caused item 4's win to evaporate, or an unguarded ingest path
 *    quietly refilled the store, no row in an append-only table would ever show
 *    it. Here, every row re-runs EVERY prior item's signal, so a win that decays
 *    is visible on the next row rather than never.
 *
 * 2. Claims could be fitted after the fact. An item that shipped against
 *    Recall@5 could be defended with "but MRR improved" once Recall@5 came in
 *    flat. A signal must be declared BEFORE the run, with its direction and
 *    magnitude, and `declaredAt` is what makes that checkable.
 *
 * The sealed dev/test split stops us fitting the DATA. This stops us fitting
 * the CLAIM. They are different failure modes and need different machinery.
 *
 * @module v1/cli/knowledge/eval/signals
 */
import type { StoreProfile } from './harness.js';
export interface PreRegisteredSignal {
    /** Stable id, referenced by the scoreboard row that shipped the item. */
    id: string;
    /** Backlog item this signal belongs to. */
    item: string;
    /** ISO date the prediction was recorded. MUST predate the run that tests it —
     *  a magnitude written down after seeing the number is not a prediction. */
    declaredAt: string;
    /** One sentence, in words, of what is claimed. */
    claim: string;
    /** Dotted path into the eval report, e.g.
     *  'results.dense-only (gte-modernbert-base).scoreboard.recallAt5'. */
    metric: string;
    direction: 'increase' | 'decrease' | 'no-worse-than';
    /** Declared before measuring. Compared against the CI half-width at scoring
     *  time: a prediction smaller than the noise floor is not testable. */
    expectedMagnitude: number;
    /**
     * Store profiles in which this signal is VISIBLE AT ALL. Item 4's superseded
     * rows do not exist in a 'fresh' corpus, so scoring it there yields a flat
     * number that means "cannot see", not "no effect". Getting this wrong is how
     * a working item gets dropped.
     */
    visibleIn: StoreProfile[];
    /** What would make this win decay. This is what populates the regression
     *  suite — declared up front, not reconstructed later from a plan. */
    decayCondition: string;
    /** Value when the item shipped. Written ONCE, never edited. */
    shipValue?: number;
    /** Value before the item shipped. Written ONCE, never edited. */
    baselineValue?: number;
    /**
     * The regime the reference value was measured under: corpus hash, golden-set
     * version and split scheme. If the current run does not match, the reference
     * describes a world that no longer exists and comparing against it is
     * meaningless — the signal reports `stale-baseline`, NOT `DECAYED`.
     *
     * This exists because it fired on its first run: the item-1 gate's 0.500
     * baseline was measured on the rank-based split, and the stability fix
     * changed DEV membership. The suite dutifully reported DECAYED for a number
     * that had not regressed at all — it had been re-measured over a different
     * set. A false alarm is not harmless: it teaches people to ignore the alarm.
     */
    measuredUnder?: {
        corpusHash?: string;
        goldenSetVersion?: string;
        splitScheme?: string;
    };
}
/**
 * Why a delta came out flat. Three different situations produce an identical
 * null row and only one of them justifies dropping an item, so a null must say
 * which it is. `cannot-see-mechanism` is the DEFAULT: claiming "no effect"
 * requires first showing the harness could have seen an effect.
 */
export type NullVerdict = 'no-effect' | 'cannot-see-mechanism' | 'redundant' | 'undetermined';
export interface SignalResult {
    id: string;
    item: string;
    claim: string;
    metric: string;
    declaredAt: string;
    currentValue: number | null;
    shipValue: number | null;
    baselineValue: number | null;
    /** Movement since ship time. Negative for a signal declared 'increase' is decay. */
    deltaSinceShip: number | null;
    verdict: 'holding' | 'DECAYED' | 'cannot-see' | 'not-yet-shipped' | 'below-noise-floor' | 'stale-baseline';
    nullVerdict?: NullVerdict;
    note: string;
}
/** Read a dotted path, tolerating keys that themselves contain dots. */
export declare function readMetric(report: unknown, dotted: string): number | null;
/**
 * THE PERSISTENT REGRESSION SUITE.
 *
 * Append one entry per shipped item, with `shipValue` and `baselineValue`
 * frozen at ship time. Never edit an existing entry: an edited prediction is
 * not a prediction. Every future scoreboard row re-scores all of them.
 */
export declare const SIGNAL_REGISTRY: PreRegisteredSignal[];
export interface Regime {
    corpusHash: string;
    goldenSetVersion: string;
    splitScheme: string;
}
export declare function scoreSignals(report: unknown, storeProfile: StoreProfile, noiseFloor: number, regime?: Regime): SignalResult[];
//# sourceMappingURL=signals.d.ts.map