@tangle-network/agent-runtime 0.107.3 → 0.107.5

package/dist/{runtime-5zM5XxLr.js → runtime-M_GIzIwe.js}

@@ -7,7 +7,7 @@ import { S as routerChatWithUsage, b as routerBrain, i as withDriverExecutor, o
 import { C as observe, O as strategyAuthorMethod, b as sample, v as refine, x as sampleThenRefine, y as runAgentic } from "./structural-rollout-IXUEplky.js";
 import { a as notifySandboxEventObserver, i as mapSandboxToolEvent, r as mapSandboxEvent, t as createSandboxToolPartState } from "./sandbox-events-DeI5xX8P.js";
 import { Bt as runAgentRounds, Dt as createExecutorRegistry, Et as createExecutor, Ht as createSandboxLineage, Kt as gateOnDeliverable, Ot as createWorktreeCliExecutor, Ut as probeSandboxCapabilities, bt as rollingDispatch, t as supervise, zt as defaultSelectWinner } from "./supervise-BiRutHS9.js";
-import { CODING_HARNESSES, InMemoryTraceStore, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, expandProfileAxes, harnessAxisOf, makeFinding as makeFinding$1, pairedBootstrap, pairedTTest, paretoFrontier, scoreKnowledgeReadiness, wilson } from "@tangle-network/agent-eval";
+import { CODING_HARNESSES, InMemoryTraceStore, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, expandProfileAxes, harnessAxisOf, makeFinding as makeFinding$1, pairedBootstrap, paretoFrontier, scoreKnowledgeReadiness, wilcoxonSignedRank, wilson } from "@tangle-network/agent-eval";
 import { heldoutSignificance, runProfileMatrix } from "@tangle-network/agent-eval/campaign";
 import { canonicalCandidateDigest, validateAgentProfileSecurity } from "@tangle-network/agent-interface";
 import { mkdir, readFile, writeFile } from "node:fs/promises";
@@ -724,7 +724,7 @@ function pairwiseSignificance(records, opts = {}) {
 			seed: 7,
 			statistic: "median"
 		});
-		const p = pairedTTest(aScores, bScores).p;
+		const test = wilcoxonSignedRank(aScores, bScores, { seed: 7 });
 		raw.push({
 			a: labelOf(ka),
 			b: labelOf(kb),
@@ -732,12 +732,16 @@ function pairwiseSignificance(records, opts = {}) {
 			delta: boot.median,
 			ciLow: boot.low,
 			ciHigh: boot.high,
-			p
+			nonZeroPairs: test.nNonZero,
+			testMethod: test.method,
+			pFloor: test.pFloor,
+			p: test.p
 		});
 	}
-	const { significant } = benjaminiHochberg(raw.map((r) => r.p), opts.fdr ?? .05);
+	const { qValues, significant } = benjaminiHochberg(raw.map((result) => result.p), opts.fdr ?? .05);
 	return raw.map((r, i) => ({
 		...r,
+		q: qValues[i] ?? 1,
 		significant: (significant[i] ?? false) && r.pairs >= minPairs
 	}));
 }
@@ -780,15 +784,18 @@ function renderLeaderboardMarkdown(report) {
 function renderPairwiseMarkdown(verdicts, title = "Pairwise significance (paired, BH-corrected)") {
 	const lines = [`## ${title}`, ""];
 	if (verdicts.length === 0) return lines.concat("_no comparable pairs_").join("\n");
-	lines.push("| A vs B | Δ(b−a) median | 95% CI | pairs | p | verdict |");
-	lines.push("|---|--:|--:|--:|--:|---|");
+	lines.push("| A vs B | Δ(b−a) median | 95% CI | pairs | changed | test | p | p floor | q | verdict |");
+	lines.push("|---|--:|--:|--:|--:|---|--:|--:|--:|---|");
 	for (const v of verdicts) {
-		const verdict = v.significant ? `**${v.delta >= 0 ? v.b : v.a} wins**` : "ns";
-		lines.push(`| ${v.a} vs ${v.b} | ${v.delta >= 0 ? "+" : ""}${pct(v.delta)} | [${pct(v.ciLow)}, ${pct(v.ciHigh)}] | ${v.pairs} | ${v.p.toFixed(3)} | ${verdict} |`);
+		const verdict = v.nonZeroPairs === 0 ? "tie" : v.significant ? `**${v.delta >= 0 ? v.b : v.a} wins**` : "ns";
+		lines.push(`| ${v.a} vs ${v.b} | ${v.delta >= 0 ? "+" : ""}${pct(v.delta)} | [${pct(v.ciLow)}, ${pct(v.ciHigh)}] | ${v.pairs} | ${v.nonZeroPairs} | ${v.testMethod} | ${formatProbability(v.p)} | ${formatProbability(v.pFloor)} | ${formatProbability(v.q)} | ${verdict} |`);
 	}
-	lines.push("", "> `ns` = not significant after Benjamini–Hochberg (or below the paired-count floor).");
+	lines.push("", "> `changed` excludes exact ties. `p` is the raw two-sided signed-rank probability; `q` is Benjamini-Hochberg adjusted. `ns` means q did not clear the target or the pair count was below the configured minimum.");
 	return lines.join("\n");
 }
+function formatProbability(value) {
+	return value > 0 && value < .001 ? value.toExponential(2) : value.toFixed(3);
+}
 function esc(s) {
 	return s.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
 }
@@ -5405,4 +5412,4 @@ function tail(s) {
 //#endregion
 export { defineLeaderboard as $, trajectoryReport as A, flatWidenGate as B, strategyAuthorContract as C, sanitizeMcpToolSchema as Ct, runBenchmark as D, secretEnvOfMcpServer as Dt, printBenchmarkReport as E, resolveSecretEnv as Et, registerShape as F, verify as G, panel as H, FileCorpus as I, buildSteerContext as J, widen as K, InMemoryCorpus as L, runPersonified as M, builtinShapes as N, promotionGate as O, createShapeRegistry as P, harvestCorpus as Q, renderCorpusToInstructions as R, authorStrategy as S, createMcpEnvironment as St, openSandboxRun as T, mcpSecretEnvMetadataKey as Tt, pipeline as U, loopUntil as V, selectValidWinner as W, registryScopeAnalyst as X, createScopeAnalyst as Y, inProcessSandboxClient as Z, discriminatingMeans as _, auditIntent as _t, localShell as a, loopCampaignDispatch as at, selectChampion as b, connectStdioMcp as bt, createVerifierEnvironment as c, deterministicCompletion as ct, worktreeFanout as d, leaderboard as dt, dumbDriver as et, analyzeTrace as f, pairwiseSignificance as ft, streamAgentTurn as g, renderPairwiseMarkdown as gt, collectAgentTurn as h, renderLeaderboardSvg as ht, jjWorkspace as i, inlineSandboxClient as it, definePersona as j, equalKOnCost as k, failuresAnalyst as l, sentinelCompletion as lt, runCoderChecks as m, renderLeaderboardMarkdown as mt, makeFinding$1 as n, resolveSandboxClient as nt, runInWorkspace as o, loopDispatch as ot, patchDelivered as p, renderLeaderboardHtml as pt, assertTraceDerivedFindings as q, gitWorkspace as r, localSandboxClient as rt, createWaterfallCollector as s, completionAuthorizes as st, computeFindingId$1 as t, naiveDriver as tt, superviseSurface as u, stopSentinel as ut, pickChampion as v, defaultAuditorInstruction as vt, SandboxRunAbortError as w, envKeyProvider as wt, assertStrategyContract as x, materializeLocalMcp as xt, runStrategyEvolution as y, McpSpawnFault as yt, fanout as z };
 
-//# sourceMappingURL=runtime-5zM5XxLr.js.map
+//# sourceMappingURL=runtime-M_GIzIwe.js.map