@tangle-network/agent-eval 0.203.2 → 0.203.4

package/dist/series-convergence-DLy0bQMV.d.ts

@@ -1,126 +0,0 @@
-import { u as JudgeScore } from "./types-DLxe96rp.js";
-import { i as ContinuousAgreementOptions, r as ContinuousAgreement } from "./judge-calibration-DFtEMlde.js";
-//#region src/statistics/agreement-irr.d.ts
-/**
- * Inter-rater reliability — Krippendorff's α under the squared-difference
- * metric, pooled across dimensions.
- *
- * Each inner array is one judge's scores. Items are matched by position
- * WITHIN a dimension: the k-th score a judge supplies carrying dimension
- * `d` is item k of `d`, and the ratings compared against each other are
- * the ones different judges gave to the same item. Every judge that scores
- * a dimension at all must supply the same number of scores for it —
- * ragged input cannot be aligned into items and throws rather than
- * comparing mismatched items.
- *
- * α = 1 − D_observed / D_expected: D_observed averages the squared
- * difference over within-item judge pairs, D_expected over every pair of
- * ratings irrespective of item. α = 1 is perfect agreement, 0 is chance,
- * negative is systematic disagreement.
- */
-declare function interRaterReliability(judgeScores: JudgeScore[][]): number;
-interface CorpusScoreRecord {
-  /** Stable identifier for the rated item (scenario, span, turn, …). */
-  itemId: string;
-  /** Identifier for the judge that produced this score. */
-  judgeName: string;
-  /** Dimension name (matches `JudgeScore.dimension`). */
-  dimension: string;
-  /** Numeric score; must be finite. */
-  score: number;
-}
-interface CorpusAgreementPerDimension extends ContinuousAgreement {
-  dimension: string;
-  /** Item IDs that contributed to this dimension's matrix (every judge scored them). */
-  itemIds: string[];
-  /** Judge IDs that contributed to this dimension's matrix. */
-  judgeIds: string[];
-}
-interface CorpusAgreementReport {
-  /** Per-dimension ICC(2,1) + κ_w + Pearson + Spearman + bootstrap CIs. */
-  perDimension: CorpusAgreementPerDimension[];
-  /** Mean ICC across dimensions (NaN if no dimension yielded a finite ICC). */
-  overallIcc: number;
-  /** Mean weighted κ across dimensions (NaN if none finite). */
-  overallWeightedKappa: number;
-  /** Dimensions evaluated (sorted). */
-  dimensions: string[];
-  /** Judges seen across the corpus (sorted). */
-  judgeIds: string[];
-}
-interface CorpusAgreementOptions extends ContinuousAgreementOptions {
-  /**
-   * Restrict the audit to these dimensions. Default = every dimension
-   * that appears in the input. A dimension named here but absent from
-   * the input throws — silent omission would corrupt the overall metric.
-   */
-  dimensions?: string[];
-  /**
-   * Restrict the audit to these judges. Default = every judge that
-   * appears in the input. A judge named here but absent from a
-   * dimension throws (see "fail loud" below).
-   */
-  judges?: string[];
-}
-/**
- * Corpus-wide inter-rater agreement across N items × M judges × D dimensions.
- *
- * For each dimension, builds the [n_items][n_judges] matrix of scores
- * (keeping only items every judge rated on that dimension), then runs
- * `continuousAgreement` to get ICC(2,1), κ_w, Pearson, Spearman, and
- * bootstrap CIs. Reports a pooled mean across dimensions as a single
- * "is this judge panel reliable on this corpus?" number.
- *
- * Fail-loud contract:
- *   - Empty input throws.
- *   - Fewer than 2 judges or fewer than 2 items per dimension throws.
- *   - A judge present in some dimensions but with zero scored items on
- *     another dimension throws (would silently shrink the matrix).
- *   - Duplicate (itemId, judgeName, dimension) records throw.
- */
-declare function corpusInterRaterAgreement(records: CorpusScoreRecord[], opts?: CorpusAgreementOptions): CorpusAgreementReport;
-/**
- * Convenience adapter for `JudgeScore[]` data keyed externally by item.
- *
- * Use when you have per-item arrays of `JudgeScore[]` (e.g. one
- * `ScenarioResult.judgeScores` per scenario) and want corpus-wide
- * agreement without manually flattening. `itemId` must be unique per
- * row of `itemsScores`.
- */
-declare function corpusInterRaterAgreementFromJudgeScores(itemsScores: Array<{
-  itemId: string;
-  scores: JudgeScore[];
-}>, opts?: CorpusAgreementOptions): CorpusAgreementReport;
-//#endregion
-//#region src/series-convergence.d.ts
-/**
- * Series convergence — detects whether a sequence of scalar measurements
- * is stabilizing, drifting, or noisy. It reads drift *across* runs, e.g.
- * "are my nightly eval scores stabilizing?".
- *
- * Three signals:
- *   - stabilized: last K values have low variance (< epsilon) — done
- *   - drifting:   recent trend is monotonic and beyond noise — regressing or improving
- *   - noisy:      neither — keep iterating, but flag as untrustworthy for gating
- */
-interface SeriesConvergenceOptions {
-  /** Window size for "recent" analysis (default 5). */
-  window?: number;
-  /** Coefficient-of-variation threshold below which the window is stabilized (default 0.05 = 5%). */
-  stableCv?: number;
-  /** Minimum monotone run length to call drift (default 3). */
-  driftRun?: number;
-}
-interface SeriesConvergenceResult {
-  state: 'stabilized' | 'drifting-up' | 'drifting-down' | 'noisy' | 'insufficient-data';
-  windowMean: number;
-  windowCv: number;
-  /** Longest monotonic run at the tail of the series (positive for up, negative for down). */
-  tailRun: number;
-  /** True when n ≥ window AND windowCv ≤ stableCv. */
-  stable: boolean;
-}
-declare function analyzeSeries(values: number[], options?: SeriesConvergenceOptions): SeriesConvergenceResult;
-//#endregion
-export { CorpusAgreementPerDimension as a, corpusInterRaterAgreement as c, CorpusAgreementOptions as i, corpusInterRaterAgreementFromJudgeScores as l, SeriesConvergenceResult as n, CorpusAgreementReport as o, analyzeSeries as r, CorpusScoreRecord as s, SeriesConvergenceOptions as t, interRaterReliability as u };
-//# sourceMappingURL=series-convergence-DLy0bQMV.d.ts.map

package/dist/series-convergence-DLy0bQMV.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"series-convergence-DLy0bQMV.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../src/series-convergence.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;;;;;;;;;;;;UCvRc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR"}