@tangle-network/agent-eval 0.200.1 → 0.202.0

package/dist/{tree-Ki_gW7z8.js → tree-D48IJUK0.js}

@@ -1,11 +1,11 @@
 import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-BsZVgBhW.js";
-import { a as medianInPlace, s as zQuantile } from "./internal-BMFSR8Ns.js";
+import { a as medianInPlace, s as zQuantile } from "./internal-BOe3zmjQ.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { r as holm } from "./multiplicity-D9cjmStq.js";
-import { a as pairedSignTest, n as DECISION_PAIRED_DELTA_STATISTIC } from "./paired-tests-C8iCsioC.js";
-import { i as pairedDeltaTest, r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-D5B1Cxc2.js";
-import { C as searchEstimateMethod, D as paretoFrontier, T as seedFromDigest, _ as rankingSplit, a as searchPolicyView, b as estimateNode, g as INSUFFICIENT_FROM, u as verifySearchClaim, v as summarizeSamples, w as searchPosterior } from "./search-kernel-AVgIjXaI.js";
-import { f as searchUnitScores } from "./search-ledger-CL0-mXQ8.js";
+import { a as pairedSignTest, n as DECISION_PAIRED_DELTA_STATISTIC } from "./paired-tests-Ccr16gcW.js";
+import { i as pairedDeltaTest, r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-DhrmjdV8.js";
+import { C as searchEstimateMethod, D as paretoFrontier, T as seedFromDigest, _ as rankingSplit, a as searchPolicyView, b as estimateNode, g as INSUFFICIENT_FROM, u as verifySearchClaim, v as summarizeSamples, w as searchPosterior } from "./search-kernel-Dpniloks.js";
+import { f as searchUnitScores } from "./search-ledger-CXQPwdWz.js";
 import { defineInlineResource } from "@tangle-network/agent-interface";
 //#region src/search/lenses/edit-credit.ts
 /**
@@ -46,29 +46,35 @@ const INTERACTION_ALPHA = .05;
 const DEFAULT_INTERACTION_GENES = 24;
 /** Signal name a policy reads. */
 const EDIT_CREDIT_SIGNAL = "reusableHunks";
+/**
+* Every choice a credit depends on. Its digest is the estimator revision each
+* credit names, so two readers that report the same revision and the same
+* `cellSetDigest` report the same numbers.
+*/
+const CREDIT_DEFINITION = {
+	name: "tangle.edit-credit.2026-09",
+	gene: "a paragraph of whitespace-normalized lines added (insert) or removed (delete) at one artifact location",
+	step: "one lineage parent and its child, both with known content and neither invalid",
+	clean: "the gene flips on the step and every gene that flips was introduced on one edge that introduced it",
+	unit: "mean of a node’s scored cells in the unit; per unit, carrier-side and lacking-side means averaged over the clean steps where both sides scored it",
+	statistic: DECISION_PAIRED_DELTA_STATISTIC,
+	interval: "percentile paired bootstrap",
+	confidence: CONFIDENCE,
+	resamples: RESAMPLES,
+	rng: "mulberry32",
+	seed: "the first 32 bits of cellSetDigest",
+	signTest: "exact, one-sided toward improvement in the objective direction",
+	methods: {
+		none: "below 2 units",
+		insufficient: `below ${DESCRIPTIVE_FROM} units`,
+		descriptive: `below 20 units`,
+		bootstrap: `from 20 units`
+	}
+};
 /** The estimator every gene credit names. */
 const EDIT_CREDIT_ESTIMATOR = {
 	uri: "npm:@tangle-network/agent-eval/search#editCredit",
-	revision: hashCanonical({
-		name: "tangle.edit-credit.2026-09",
-		gene: "a paragraph of whitespace-normalized lines added (insert) or removed (delete) at one artifact location",
-		step: "one lineage parent and its child, both with known content and neither invalid",
-		clean: "the gene flips on the step and every gene that flips was introduced on one edge that introduced it",
-		unit: "mean of a node’s scored cells in the unit; per unit, carrier-side and lacking-side means averaged over the clean steps where both sides scored it",
-		statistic: DECISION_PAIRED_DELTA_STATISTIC,
-		interval: "percentile paired bootstrap",
-		confidence: CONFIDENCE,
-		resamples: RESAMPLES,
-		rng: "mulberry32",
-		seed: "the first 32 bits of cellSetDigest",
-		signTest: "exact, one-sided toward improvement in the objective direction",
-		methods: {
-			none: "below 2 units",
-			insufficient: `below ${DESCRIPTIVE_FROM} units`,
-			descriptive: `below 20 units`,
-			bootstrap: `from 20 units`
-		}
-	})
+	revision: hashCanonical(CREDIT_DEFINITION)
 };
 /**
 * The edit-credit lens. Pure: the result depends only on `state` and what
@@ -350,13 +356,13 @@ function editCredit(state, options) {
 			chance: {
 				measuredEdits: measuredLeads.length,
 				expectedReusable,
-				method: `per measured edit, the chance with no true effect that its verdict is reusable: the exact one-sided sign test's attained size at its unit count below 20 units (at most ${(1 - CONFIDENCE) / 2}; the point-estimate condition only lowers it), the nominal ${(1 - CONFIDENCE) / 2} of the percentile bootstrap from 20; summed, so it is the expected number of false reusable edits if every edit had no effect`
+				method: `per measured edit, the chance with no true effect that its verdict is reusable: the exact one-sided sign test's attained size at its unit count below 20 units (at most ${.050000000000000044 / 2}; the point-estimate condition only lowers it), the nominal ${.050000000000000044 / 2} of the percentile bootstrap from 20; summed, so it is the expected number of false reusable edits if every edit had no effect`
 			},
 			lineage,
 			method: {
 				estimator: EDIT_CREDIT_ESTIMATOR,
 				credit: `per clean step (a parent and its child on which every change belongs to one edit that introduced the gene), the carrying side against the lacking side on each unit both scored; per unit, both sides averaged over those steps; ${DECISION_PAIRED_DELTA_STATISTIC} paired delta with a ${Math.round(CONFIDENCE * 100)}% percentile bootstrap (${RESAMPLES} resamples, seeded from cellSetDigest): none below 2 units, insufficient below ${DESCRIPTIVE_FROM}, a descriptive interval and exact one-sided sign p below 20, a decision-grade interval from 20; unscored cells are absent, never zero`,
-				verdict: `pairedDeltaTest's own decision in each direction: reusable when it finds an improvement (below 20 units an exact one-sided sign test at ${(1 - CONFIDENCE) / 2} with the point estimate on the better side; from 20 the bootstrap interval above zero), harmful when it finds one in the other direction, unresolved otherwise, insufficient below ${DESCRIPTIVE_FROM} units`,
+				verdict: `pairedDeltaTest's own decision in each direction: reusable when it finds an improvement (below 20 units an exact one-sided sign test at ${.050000000000000044 / 2} with the point estimate on the better side; from 20 the bootstrap interval above zero), harmful when it finds one in the other direction, unresolved otherwise, insufficient below ${DESCRIPTIVE_FROM} units`,
 				interaction: `per unit, a gene's clean-step contrast where both ends carry the other gene minus where neither does, averaged over both directions; ${Math.round(CONFIDENCE * 100)}% percentile bootstrap (${RESAMPLES} resamples) for spread; exact two-sided sign test, Holm-adjusted over the pairs tested, flagged at ${INTERACTION_ALPHA}`,
 				multiplicity: "credit is not adjusted for the number of genes: it steers which edits to reuse and test next, as selection estimates steer spend, and claims nothing; interaction flags are Holm-adjusted because pairs grow with the square of the genes",
 				selectionBias: "a policy chooses a parent for its scores, so a parent is high by chance on the units it was chosen on: on a step where the child gained a gene the lacking side is that parent, and credit is biased toward harmful; where the child lost it, toward reusable. Only the sealed test split is free of this, so a reusable edit still needs a held-out test before it ships"
@@ -644,7 +650,7 @@ function pairedEstimate(input) {
 * smallest positive count that clears α) below the bootstrap gate, the
 * percentile bootstrap's nominal one-sided α from it. */
 const chanceReusable = (() => {
-	const alpha = (1 - CONFIDENCE) / 2;
+	const alpha = .050000000000000044 / 2;
 	const cache = /* @__PURE__ */ new Map();
 	return (pairs) => {
 		if (pairs < DESCRIPTIVE_FROM) return 0;
@@ -653,7 +659,8 @@ const chanceReusable = (() => {
 		if (cached !== void 0) return cached;
 		let size = 0;
 		for (let positive = 0; positive <= pairs; positive++) {
-			const p = pairedSignTest(Array.from({ length: pairs }, (_, index) => index < positive ? 1 : -1), "greater").pValue;
+			const signs = Array.from({ length: pairs }, (_, index) => index < positive ? 1 : -1);
+			const p = pairedSignTest(signs, "greater").pValue;
 			if (p <= alpha) {
 				size = p;
 				break;
@@ -785,7 +792,8 @@ function interactionEstimate(genes, steps, values, sign) {
 		signP: null,
 		synergy: sign * interaction > 0
 	};
-	const test = pairedDeltaTest(values.map(() => 0), values, {
+	const zeros = values.map(() => 0);
+	const test = pairedDeltaTest(zeros, values, {
 		statistic: DECISION_PAIRED_DELTA_STATISTIC,
 		confidence: CONFIDENCE,
 		resamples: RESAMPLES,
@@ -4096,4 +4104,4 @@ function tree(state) {
 //#endregion
 export { editCreditText as A, draftOnPlateau as C, EDIT_CREDIT_ESTIMATOR as D, front as E, EDIT_CREDIT_SIGNAL as O, vectorEmbedding as S, screenedNodes as T, lineageEdits as _, skillCalibration as a, surfaceDigestEdits as b, META_SEARCH_SIGNAL as c, objectiveKey as d, renderMetaSearchText as f, lineEditDistance as g, landscape as h, nextUnitExtension as i, editCredit as k, metaSearch as l, formatLandscape as m, taskMatrix as n, skillManifold as o, searchPolicyGenome as p, formatSkillManifold as r, META_SEARCH_SCORE_SOURCE as s, tree as t, metaSearchScore as u, nearestNeighbours as v, searchPlateau as w, surfaceTextEdits as x, profileTextLines as y };
 
-//# sourceMappingURL=tree-Ki_gW7z8.js.map
+//# sourceMappingURL=tree-D48IJUK0.js.map