dsh-llm-verifier 0.1.0

package/lib/client.js

@@ -0,0 +1,385 @@
+window.__ModuleLoader__.load({
+	id: "dsh-llm-verifier",
+	factory: (require) => {
+		var module = { exports: {} };
+		var exports = module.exports;
+		Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+		let _deepseek_ai_dsh_client_ui_primitives = require("@deepseek-ai/dsh-client-ui-primitives");
+		let react = require("react");
+		let react_jsx_runtime = require("react/jsx-runtime");
+		//#region src/client.tsx
+		const NS = "llm-verifier";
+		const shell = {
+			display: "flex",
+			flexDirection: "column",
+			gap: 18,
+			padding: "8px 4px 32px",
+			color: "var(--dsw-text-primary)"
+		};
+		const card = {
+			display: "flex",
+			flexDirection: "column",
+			gap: 0,
+			padding: "16px 16px 0",
+			border: "1px solid var(--dsw-alias-border-l2, rgba(255, 255, 255, 0.16))",
+			borderRadius: 12,
+			background: "var(--dsw-alias-bg-module, rgba(20, 31, 57, 0.42))",
+			overflow: "hidden"
+		};
+		const sectionTitle = {
+			display: "flex",
+			gap: 10,
+			alignItems: "center",
+			padding: "0 0 12px",
+			borderBottom: "1px solid var(--dsw-alias-border-l2, rgba(255, 255, 255, 0.16))"
+		};
+		const row = {
+			display: "grid",
+			gridTemplateColumns: "minmax(150px, 1fr) minmax(220px, 1.4fr)",
+			gap: 18,
+			alignItems: "center",
+			padding: "14px 0",
+			borderBottom: "1px solid var(--dsw-alias-border-l2, rgba(255, 255, 255, 0.16))"
+		};
+		const selectStyle = {
+			width: "100%",
+			minHeight: 38,
+			padding: "0 12px",
+			borderRadius: 10,
+			color: "var(--dsw-text-primary)",
+			background: "var(--dsw-surface-sunken)",
+			border: "1px solid var(--dsw-alias-border-l2, rgba(255, 255, 255, 0.16))"
+		};
+		function record(value) {
+			return typeof value === "object" && value !== null && !Array.isArray(value) ? value : {};
+		}
+		function values(view) {
+			const v = record(view.value);
+			return {
+				provider: String(v.provider ?? ""),
+				model: String(v.model ?? ""),
+				...typeof v.reasoningEffort === "string" ? { reasoningEffort: v.reasoningEffort } : {},
+				maxTokens: Number(v.maxTokens ?? 32768),
+				maxConcurrency: Number(v.maxConcurrency ?? 8),
+				maxRetries: Number(v.maxRetries ?? 3),
+				timeoutMs: Number(v.timeoutMs ?? 3e5),
+				cacheMaxEntries: Number(v.cacheMaxEntries ?? 1e4),
+				estimatedInputUsdPerMillion: Number(v.estimatedInputUsdPerMillion ?? 0),
+				estimatedOutputUsdPerMillion: Number(v.estimatedOutputUsdPerMillion ?? 0)
+			};
+		}
+		function message(error) {
+			return error instanceof Error ? error.message : String(error);
+		}
+		function Label({ title, help }) {
+			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+				style: { fontWeight: 600 },
+				children: title
+			}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+				style: {
+					fontSize: 12,
+					color: "var(--dsw-text-secondary)",
+					marginTop: 3
+				},
+				children: help
+			})] });
+		}
+		function VerifierSettings({ api }) {
+			const [loaded, setLoaded] = (0, react.useState)(null);
+			const [draft, setDraft] = (0, react.useState)(null);
+			const [busy, setBusy] = (0, react.useState)(false);
+			const [error, setError] = (0, react.useState)(null);
+			const [saved, setSaved] = (0, react.useState)(false);
+			const load = async () => {
+				setError(null);
+				try {
+					const [m, s] = await Promise.all([api.llm.models({}), api.settings.describe({})]);
+					if (!m.result.ok) throw new Error(m.result.error.message);
+					if (!s.result.ok) throw new Error(s.result.error.message);
+					const view = s.result.value.namespaces.find((x) => x.ns === NS);
+					if (!view) throw new Error("Verifier settings namespace is not registered. Restart the DSH host.");
+					const next = {
+						groups: m.result.value.groups,
+						settings: view,
+						writable: s.result.value.writable,
+						failures: m.result.value.failures.map((f) => f.name + ": " + f.message)
+					};
+					setLoaded(next);
+					setDraft(values(view));
+				} catch (e) {
+					setError(message(e));
+				}
+			};
+			(0, react.useEffect)(() => {
+				load();
+			}, []);
+			const models = (0, react.useMemo)(() => loaded?.groups.find((g) => g.id === draft?.provider)?.models ?? [], [loaded, draft?.provider]);
+			const efforts = models.find((m) => m.id === draft?.model)?.reasoning?.efforts ?? [];
+			const patch = (key, value) => setDraft((v) => v ? {
+				...v,
+				[key]: value
+			} : v);
+			const save = async () => {
+				if (!loaded || !draft) return;
+				setBusy(true);
+				setSaved(false);
+				setError(null);
+				try {
+					const section = {
+						...record(loaded.settings.user),
+						...draft
+					};
+					if (!draft.reasoningEffort) delete section.reasoningEffort;
+					const res = await api.settings.update({
+						ns: NS,
+						patch: section,
+						expectedRevision: loaded.settings.revision
+					});
+					if (!res.result.ok) throw new Error(res.result.error.message);
+					setLoaded((v) => v ? {
+						...v,
+						settings: res.result.value
+					} : v);
+					setDraft(values(res.result.value));
+					setSaved(true);
+				} catch (e) {
+					setError(message(e));
+				} finally {
+					setBusy(false);
+				}
+			};
+			if (!loaded || !draft) return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+				style: shell,
+				children: [
+					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("h2", { children: "LLM Verifier" }),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", { children: error ?? "正在读取 DSH 模型和设置…" }),
+					error && /* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Button, {
+						onClick: () => void load(),
+						children: "重试"
+					})
+				]
+			});
+			const numeric = (key, min = 0) => /* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Input, {
+				type: "number",
+				min,
+				value: String(draft[key]),
+				onChange: (e) => patch(key, Number(e.target.value))
+			});
+			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+				style: shell,
+				children: [
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("h2", {
+						style: { margin: "0 0 6px" },
+						children: "LLM Verifier"
+					}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+						style: {
+							margin: 0,
+							color: "var(--dsw-text-secondary)"
+						},
+						children: "选择任意已在 DSH「模型」页配置并启用的模型作为独立裁判。设置实时生效。"
+					})] }),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+						style: card,
+						children: [
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: sectionTitle,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.StateDot, { state: "done" }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: "裁判模型" })]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "供应商",
+									help: "只显示当前 DSH 中可路由的供应商"
+								}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("select", {
+									style: selectStyle,
+									value: draft.provider,
+									onChange: (e) => {
+										const provider = e.target.value;
+										const first = loaded.groups.find((g) => g.id === provider)?.models[0];
+										setDraft({
+											...draft,
+											provider,
+											...first ? {
+												model: first.id,
+												reasoningEffort: first.reasoning?.defaultEffort
+											} : {}
+										});
+									},
+									children: loaded.groups.map((g) => /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("option", {
+										value: g.id,
+										children: [
+											g.name,
+											" · ",
+											g.id
+										]
+									}, g.id))
+								})]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "模型",
+									help: "模型目录来自 DSH adapter，选择结果会持久化"
+								}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("select", {
+									style: selectStyle,
+									value: draft.model,
+									onChange: (e) => {
+										const model = e.target.value;
+										const found = models.find((m) => m.id === model);
+										setDraft({
+											...draft,
+											model,
+											...found?.reasoning?.defaultEffort ? { reasoningEffort: found.reasoning.defaultEffort } : { reasoningEffort: void 0 }
+										});
+									},
+									children: models.map((m) => /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("option", {
+										value: m.id,
+										children: [
+											m.name,
+											" · ",
+											m.id
+										]
+									}, m.id))
+								})]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "推理强度",
+									help: "由所选模型 adapter 声明；留空使用模型默认值"
+								}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("select", {
+									style: selectStyle,
+									value: draft.reasoningEffort ?? "",
+									onChange: (e) => patch("reasoningEffort", e.target.value || void 0),
+									children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+										value: "",
+										children: "模型默认"
+									}), efforts.map((e) => /* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+										value: e.id,
+										children: e.name
+									}, e.id))]
+								})]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "最大输出 Token",
+									help: "每个裁判请求的输出上限"
+								}), numeric("maxTokens", 1)]
+							})
+						]
+					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+						style: card,
+						children: [
+							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+								style: sectionTitle,
+								children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: "执行控制" })
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "最大并发",
+									help: "所有 verifier 工具共享的请求并发上限"
+								}), numeric("maxConcurrency", 1)]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "最多重试",
+									help: "短暂网络、限流和服务端错误的重试次数"
+								}), numeric("maxRetries", 0)]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "请求超时（毫秒）",
+									help: "单个模型请求的超时时间"
+								}), numeric("timeoutMs", 1)]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "缓存条目上限",
+									help: "持久评分缓存保留的最大条目数"
+								}), numeric("cacheMaxEntries", 1)]
+							})
+						]
+					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+						style: card,
+						children: [
+							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+								style: sectionTitle,
+								children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: "费用估算（每百万 Token，USD）" })
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "输入价格",
+									help: "仅用于结果中的 estimatedCostUsd"
+								}), numeric("estimatedInputUsdPerMillion", 0)]
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+								style: row,
+								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(Label, {
+									title: "输出价格",
+									help: "仅用于结果中的 estimatedCostUsd"
+								}), numeric("estimatedOutputUsdPerMillion", 0)]
+							})
+						]
+					}),
+					loaded.failures.length > 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+						style: {
+							...card,
+							borderColor: "var(--dsw-alias-state-warn-primary, #d9a441)",
+							paddingBottom: 16
+						},
+						children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: "部分模型目录读取失败" }), loaded.failures.map((x) => /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", { children: x }, x))]
+					}),
+					error && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+						style: { color: "var(--dsw-danger)" },
+						children: error
+					}),
+					saved && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("div", {
+						style: { color: "var(--dsw-success)" },
+						children: "已保存，后续 verifier 调用将使用新设置。"
+					}),
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+						style: {
+							display: "flex",
+							gap: 10
+						},
+						children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Button, {
+							disabled: busy || !loaded.writable,
+							onClick: () => void save(),
+							children: busy ? "保存中…" : "保存设置"
+						}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)(_deepseek_ai_dsh_client_ui_primitives.Button, {
+							variant: "outline",
+							disabled: busy,
+							onClick: () => void load(),
+							children: "重新载入"
+						})]
+					})
+				]
+			});
+		}
+		const inject = ["slots", "connection"];
+		function apply(ctx) {
+			const connection = ctx.get("connection");
+			ctx.slots.inject("settings.section", () => ctx.slots.register({
+				name: "settings.section",
+				id: "llm-verifier",
+				order: 35,
+				label: "LLM Verifier",
+				inject: () => ({ api: connection.api })
+			}, VerifierSettings));
+		}
+		//#endregion
+		exports.apply = apply;
+		exports.inject = inject;
+		return module.exports;
+	}
+});
+
+//# sourceMappingURL=client.js.map

package/lib/core.js

@@ -0,0 +1,163 @@
+//#region src/core.ts
+const GRANULARITY = 20;
+const LETTERS = Array.from({ length: 20 }, (_, index) => String.fromCharCode(65 + index));
+const SCALE_DESCRIPTION = [
+	"Rate how likely the agent correctly solved the task on a 20-point scale using letters A through T:",
+	"  A = clearly and completely succeeded with verified output (best)",
+	"  B-D = succeeded with only minor issues",
+	"  E-G = above average, mostly correct with some issues",
+	"  H-J = uncertain, leans toward success",
+	"  K-M = uncertain, leans toward failure",
+	"  N-P = below average, significant issues remain",
+	"  Q-S = failed with some partial progress",
+	"  T = clearly and completely failed (worst)"
+].join("\n");
+const DEFAULT_CRITERIA = [
+	{
+		id: "specification",
+		name: "Specification Adherence",
+		description: "Re-read the task description and check exact requirements: file paths, output formats, naming, and explicit constraints. Penalize a solution that solves a similar but different problem."
+	},
+	{
+		id: "output_match",
+		name: "Output Match",
+		description: "Find the final verification command and compare its actual stdout/stderr to the required output. Reward only evidence literally visible in observed output; do not trust narration."
+	},
+	{
+		id: "error_signals",
+		name: "Error Signal Detection",
+		description: "Scan especially later steps for unresolved errors, tracebacks, non-zero exits, command-not-found, missing files, compilation failures, and test failures. Score only unresolved error evidence."
+	}
+];
+const DEFAULT_GROUND_TRUTH_NOTE = "**IMPORTANT:** Focus on observed tool and terminal output as ground truth. Do NOT trust the agent's self-assessment or claims of success.";
+function normalizeScoreLetter(token) {
+	let value = token.trim();
+	if (value.startsWith(">")) value = value.slice(1).trim();
+	return /^([A-T])$/i.exec(value)?.[1]?.toUpperCase();
+}
+function letterValue(letter) {
+	return 20 - (letter.charCodeAt(0) - 65);
+}
+function findTagLogprobs(tokens, positions, tag) {
+	if (tokens.length === 0 || positions.length === 0) return void 0;
+	for (const suffix of [tag, tag.slice(0, -1)]) {
+		let found;
+		let text = "";
+		for (let index = 0; index < tokens.length; index += 1) {
+			text += tokens[index];
+			if (text.trimEnd().endsWith(suffix) && index + 1 < positions.length) found = positions[index + 1];
+		}
+		if (found !== void 0) return found;
+	}
+}
+function extractScore(completion, tag) {
+	const alternatives = findTagLogprobs(completion.tokens, completion.positions, tag);
+	const probabilities = /* @__PURE__ */ new Map();
+	for (const alternative of alternatives ?? []) {
+		const letter = normalizeScoreLetter(alternative.token);
+		if (letter === void 0 || !Number.isFinite(alternative.logprob)) continue;
+		const value = letterValue(letter);
+		probabilities.set(value, Math.max(probabilities.get(value) ?? 0, Math.exp(alternative.logprob)));
+	}
+	if (probabilities.size > 0) {
+		let probability = 0;
+		let expectation = 0;
+		for (const [value, p] of probabilities) {
+			probability += p;
+			expectation += value * p;
+		}
+		if (probability > 0) return (expectation / probability - 1) / 19;
+	}
+	const name = tag.slice(1, -1);
+	const regex = new RegExp("<" + name + ">\\s*(.+?)\\s*</" + name + ">", "gi");
+	let last = null;
+	for (let match = regex.exec(completion.text); match !== null; match = regex.exec(completion.text)) last = match;
+	const letter = normalizeScoreLetter(last?.[1] ?? "");
+	return letter === void 0 ? .5 : (letterValue(letter) - 1) / 19;
+}
+function buildPairwisePrompt(problem, traceA, traceB, criterion, groundTruthNote = DEFAULT_GROUND_TRUTH_NOTE) {
+	return [
+		"You are an expert evaluator of AI coding agents. You will see a task description and two agent trajectories, then evaluate them on ONE specific criterion, stated at the end.",
+		groundTruthNote,
+		"**Task:**\n" + problem,
+		"**Trajectory A:**\n" + traceA,
+		"**Trajectory B:**\n" + traceB,
+		"**Rating Scale:**\n" + SCALE_DESCRIPTION,
+		"**Evaluation Guideline — " + criterion.name + ":**\n" + criterion.description,
+		"Score each trajectory ONLY on this specific criterion (\"" + criterion.name + "\"). Ignore other aspects that are not relevant to it.",
+		"Reason it through first, then END your reply with exactly these two lines and nothing after them. Replace each placeholder with a single letter A-T, keeping the spaces around the letter exactly as shown:\n<score_A> LETTER_A_TO_T </score_A>\n<score_B> LETTER_A_TO_T </score_B>",
+		"Begin your analysis now."
+	].join("\n\n");
+}
+function buildProgressPrompt(problem, steps, checkpoints) {
+	const trajectory = steps.map((step, index) => "=== Agent Step " + (index + 1) + " ===\n" + step.trim()).join("\n\n");
+	const tags = checkpoints.map((_, index) => "<c" + (index + 1) + ">LETTER</c" + (index + 1) + ">").join("\n");
+	return [
+		"You are a strict, skeptical evaluator of agent task attempts. Trust observed output — NOT the agent's narration.",
+		"**Task instruction:**\n" + problem.trim(),
+		"**Agent trajectory (" + steps.length + " agent steps):**\n" + trajectory,
+		"Score whether the agent's CURRENT state at each checkpoint would actually satisfy the task's hidden grader.",
+		"Use A through T where A = certainly NO, H-M = uncertain, N-S = leans YES, and T = essentially certain YES with matching observed verification.",
+		"Effort and confident narration are not evidence. A state without real verification should not exceed K. Scores may decrease after regressions.",
+		"The checkpoints are:\n" + checkpoints.map((step, index) => "  Checkpoint " + (index + 1) + " = state right after Agent Step " + step).join("\n"),
+		"Output EXACTLY these lines and nothing else:\n" + tags
+	].join("\n\n");
+}
+/** Progress uses A=NO..T=YES, the reverse of pairwise success scoring. */
+function extractProgressScore(completion, tag) {
+	return 1 - extractScore(completion, tag);
+}
+function bradleyTerry(rewardA, rewardB) {
+	return 1 / (1 + Math.exp(-(rewardA - rewardB)));
+}
+function seededRandom(seed) {
+	let state = seed >>> 0;
+	return () => {
+		state = state + 1831565813 >>> 0;
+		let value = state;
+		value = Math.imul(value ^ value >>> 15, value | 1);
+		value ^= value + Math.imul(value ^ value >>> 7, value | 61);
+		return ((value ^ value >>> 14) >>> 0) / 4294967296;
+	};
+}
+function ringCycle(count, seed = 0) {
+	if (count <= 1) return [];
+	const permutation = Array.from({ length: count }, (_, index) => index);
+	const random = seededRandom(seed);
+	for (let index = count - 1; index > 0; index -= 1) {
+		const other = Math.floor(random() * (index + 1));
+		[permutation[index], permutation[other]] = [permutation[other], permutation[index]];
+	}
+	return permutation.map((candidate, index) => [candidate, permutation[(index + 1) % count]]);
+}
+function pivotRoundPairs(count, pivots) {
+	const pivotSet = new Set(pivots);
+	const pairs = [];
+	for (let candidate = 0; candidate < count; candidate += 1) if (!pivotSet.has(candidate)) for (const pivot of pivots) pairs.push([candidate, pivot]);
+	const sorted = [...pivots].sort((a, b) => a - b);
+	for (let left = 0; left < sorted.length; left += 1) for (let right = left + 1; right < sorted.length; right += 1) pairs.push([sorted[left], sorted[right]]);
+	return pairs;
+}
+function accumulatePairs(pairs, rewards, wins, counts) {
+	for (const [a, b] of pairs) {
+		const reward = rewards.get(a + "," + b) ?? [.5, .5];
+		const probability = bradleyTerry(reward[0], reward[1]);
+		wins[a] = (wins[a] ?? 0) + probability;
+		counts[a] = (counts[a] ?? 0) + 1;
+		wins[b] = (wins[b] ?? 0) + 1 - probability;
+		counts[b] = (counts[b] ?? 0) + 1;
+	}
+}
+function topPivots(wins, counts, requested) {
+	return Array.from({ length: wins.length }, (_, index) => index).sort((a, b) => (wins[b] ?? 0) / (counts[b] || 1) - (wins[a] ?? 0) / (counts[a] || 1) || a - b).slice(0, Math.min(requested, wins.length));
+}
+function rankScores(wins, counts) {
+	return Array.from({ length: wins.length }, (_, index) => ({
+		index,
+		score: (wins[index] ?? 0) / (counts[index] || 1)
+	})).sort((a, b) => b.score - a.score || a.index - b.index);
+}
+//#endregion
+export { DEFAULT_CRITERIA, DEFAULT_GROUND_TRUTH_NOTE, GRANULARITY, LETTERS, SCALE_DESCRIPTION, accumulatePairs, bradleyTerry, buildPairwisePrompt, buildProgressPrompt, extractProgressScore, extractScore, normalizeScoreLetter, pivotRoundPairs, rankScores, ringCycle, seededRandom, topPivots };
+
+//# sourceMappingURL=core.js.map