QS 声誉调查 · 数据可信与合规体系 QS Reputation Surveys · Data Integrity & Compliance Framework

QS 声誉调查可信核验体系 — 核验标准 · 身份核对 · 环境检测 · 内容与模式分析 QS Reputation Verification Framework — Standards · Identity Checks · Environment Consistency · Content & Pattern Analysis

26项items
QS 官方明文监测标准QS-published monitoring standards
156项checks
作答环境检测全集 · 12 组Response-environment checks · 12 groups
15项checks
身份核验 — 学者 8 项 · 雇主 7 项Identity checks — Academics 8 · Employers 7
5年years
计分池与回溯跨度Scoring pool & retrospective window
阅读方式How to read 7 块框架 · 点标题行展开细节 · 出处等级:QS公开 / 行业惯例 / 合理推断 Seven blocks · Expand a title row for detail · Source grades: QS-published / Industry practice / Inference
标准检查清单Standard checklist 7 板块 · 共 52 项 — 每项均为标准服务范围内的固定动作,逐项预勾;点开任一板块,清单在第二层最上方 7 blocks · 52 items — every item a fixed action within standard scope, pre-checked; expand any block to see its checklist at the top of the detail layer
1

QS 如何核验一份回答How QS Verifies a Response

四层体系:资格 · 真实 · 统计 · 复核Four layers: Eligibility · Credibility · Statistics · Review 26 项明文标准26 published standards阶梯处置Escalating consequences5 年计分池5-year scoring pool检查点 6 项6 checkpoints ▾
☑ 标准检查清单 · 6 项标准服务范围内的固定动作
  • ☑逐项对照 26 项 QS 明文监测标准
  • ☑核查名单与回答的来源可追溯记录
  • ☑排查激励、请托与指导作答痕迹
  • ☑复核 5 年计分池内的历史回答风险
  • ☑预警机构层提名量异常跃升
  • ☑确认处置阶梯各档风险点已排查
QS公开QS 官方文件明文 行业惯例行业普遍做法 / 可信媒体与学术来源 合理推断基于公开信息的推演,非 QS 事实

四层体系(推演结构)

资格门 · 硬筛→真实性 · 多信号加权→机构层统计异常→处置阶梯

QS 每年处理数十万份回答,人工只能看算法标出的异常样本;它能用的也只有可机读、可外部核验的数据。真实形态大概率是多信号加权的风险评估,而非单一硬线——具体阈值 QS 从不公开,公开一条线,造假者就会贴着线做。合理推断

处置阶梯

降权→剔除单份回答→整批取消→整校排除

第一组 · 作答时采集的技术信号

  • IP 地址被采集并用于验证回答QS公开
  • 由 IP 推导的大致位置(经纬度)同样用于验证QS公开
  • 设备与浏览器配置被采集用于验证QS公开
  • 验证目的明文包括:检测重复、欺诈与被操纵的回答QS公开
  • QS 记录每位受访者的信息是如何到达 QS 的(来源可追溯)QS公开
  • 回答仅 QS 团队及其平台服务商可见QS公开

第二组 · 身份与资格信号

  • 学者侧收集 ORCID;雇主侧另收集行业、组织规模、LinkedIn、组织网站等字段(收集本身为明文,是否逐项交叉核验未公开)QS公开
  • 学术指标称受访者具有「已核实的机构隶属」,但核实方法不公开QS公开
  • 雇主调查设资格 + 验证双关卡;高校与研究机构人员不具雇主受访资格,其提名被排除QS公开
  • 提名名单有邮箱与身份门槛:不用通用邮箱、不用个人邮箱服务、职位须可核实,优先机构邮箱QS公开
  • 资格定义明确:学者须从事教学或研究;雇主须实际参与招聘决策QS公开
  • 公开注册(sign-up)者逐一筛查有效性与真实性,并须注明从何处得知调查QS公开

第三组 · 规则性剔除与调查设计

  • 自投排除:学者在问卷中无法选择自己所属机构QS公开
  • 熟悉度过滤:须先声明熟悉的学科与地域,声明之外的评价被过滤或降权QS公开
  • 无效、不完整、重复、欺诈的提交被排除QS公开
  • 提名结构设上限:每人国内至多 10 所、国际至多 30 所,QS 称此结构降低扭曲可能QS公开
  • 调查由 QS 直接发送(1 封邀请 + 2 封提醒),受访者独立完成QS公开
  • 5 年计分池:近 3 年按 100%、第 4 年 50%、第 5 年 25% 计入,并以统计变换压制异常值QS公开
  • 名单两年有效,每校每项调查提名封顶 400 人QS公开

第四组 · 统计与行为检测

  • 异常检测算法盯两件事:某校表现的异常跃升与非典型回答模式;不合参数的回答被移除,异常机构被深入审查(观察名单机制)QS公开
  • 若一批回答的来源无法核实为完全独立,该调查当年全部相关回答可被取消资格QS公开
  • QS 称其验证系统会主动监测、复核并排除异常回答与模式QS公开
  • 过去两年引入了机器学习驱动的新验证技术(特征与阈值不公开)QS公开
  • 更复杂的数据科学验证会被回溯应用到全部 5 年回答——当年通过的回答,日后仍可能被重判QS公开
  • 媒体评论提到以 AI 检测「协同 / 操纵」式回答;方向与 QS 官方表述一致,但该措辞无 QS 原文出处行业惯例

一个关键判断:这 26 项里,技术信号只是最小的一块;体系的主体在跨人聚合、资格门槛、机构层统计与故意不公开的规则。这也是为什么「把自己这一份填得像真的」从来不是重点——重点是整批人的真实与独立。

相关案例

  • KAIST(韩国,2025) — 一个系以 100 美元激励海外学者参与 QS 声誉调查,被 QS 排除出 2026 年排名一期,涉事回复从未计入。 QS 官方总结 ↗
  • 科克大学 UCC(爱尔兰,2013) — 校长动员教师各招募 3 名海外学者注册 QS 投票;QS 未罚该校,但就此修改规则、废弃此类机构注册。 Inside Higher Ed ↗
  • Op4G / Slice MR(美国) — 以付费假受访者、VPN 与筛选题辅导伪造问卷数据逾 1,000 万美元,多人已认罪——跨行业参照。 美国司法部 ↗
☑ Standard checklist · 6 itemsFixed actions within standard scope
  • ☑Work through all 26 QS-published monitoring standards
  • ☑Verify provenance records for lists and responses
  • ☑Screen for incentives, solicitation and coaching traces
  • ☑Review historical-response risk inside the 5-year pool
  • ☑Flag unusual institution-level nomination jumps
  • ☑Confirm each escalation-ladder risk point has been checked
QS-publishedstated in QS official documents Industry practicestandard practice / credible media & academic sources Inferencereasoned from public information — not a QS fact

The four layers (inferred structure)

Eligibility gate→Credibility scoring→Institution-level anomaly checks→Escalating consequences

QS processes hundreds of thousands of responses a year; humans can only review what algorithms flag, and only machine-readable, externally verifiable data can feed the system. The real design is most likely a multi-signal weighted risk assessment rather than any single hard threshold — and thresholds are never published, because a published line is a line manipulators work right up against.Inference

Escalation ladder

Down-weighting→Single response removed→Whole batch disqualified→Institution excluded

Group 1 · Technical signals collected at response time

  • IP address is collected and used to verify responsesQS-published
  • Approximate location derived from the IP (coordinates) is likewise used for verificationQS-published
  • Device and browser configuration is collected for verificationQS-published
  • Stated purposes include detecting duplicate, fraudulent and manipulated responsesQS-published
  • QS records how each respondent's details reached QS — provenance is traceableQS-published
  • Responses are visible only to the QS team and its platform providersQS-published

Group 2 · Identity and eligibility signals

  • Academics are asked for an ORCID; employers additionally for industry, organisation size, LinkedIn and organisation website (collection is published; whether each field is cross-checked is not)QS-published
  • Academic indicators describe respondents as having “verified affiliations”, without publishing the verification methodQS-published
  • The employer survey applies a dual eligibility + verification gate; staff of universities and research institutions are not eligible as employers and their nominations are excludedQS-published
  • Nomination lists face email and identity thresholds: no generic mailboxes, no personal email services, roles must be verifiable, institutional email preferredQS-published
  • Eligibility is defined plainly: academics must teach or research; employers must actually take part in hiring decisionsQS-published
  • Open sign-ups are screened individually for validity and authenticity, and must state how the respondent heard of the surveyQS-published

Group 3 · Rule-based exclusions and survey design

  • Self-nomination excluded: academics cannot select their own institutionQS-published
  • Familiarity filter: respondents first declare subject and regional familiarity; ratings outside it are filtered or down-weightedQS-published
  • Invalid, incomplete, duplicate or fraudulent submissions are excludedQS-published
  • Nomination caps: at most 10 domestic and 30 international institutions per respondent — a structure QS says reduces distortionQS-published
  • The survey is sent by QS directly (1 invitation + 2 reminders) and completed independentlyQS-published
  • 5-year scoring pool: the last 3 years count at 100%, year 4 at 50%, year 5 at 25%, with statistical transforms suppressing outliersQS-published
  • Lists stay valid for two years; nominations are capped at 400 per institution per surveyQS-published

Group 4 · Statistical and behavioural detection

  • Anomaly algorithms watch for unusual jumps in an institution's performance and atypical response patterns; out-of-parameter responses are removed and anomalous institutions reviewed in depth (a watch-list mechanism)QS-published
  • If a set of responses cannot be verified as fully independent in origin, all related responses for that survey year may be disqualifiedQS-published
  • QS states its validation system actively monitors, reviews and removes anomalous responses and patternsQS-published
  • Machine-learning-driven validation has been introduced over the past two years (features and thresholds unpublished)QS-published
  • More sophisticated data-science validation is applied retrospectively to all five years of responses — a response accepted in its year can still be re-judged laterQS-published
  • Media commentary describes AI detection of “coordinated / manipulated” responses; the direction matches QS statements, but the wording has no QS sourceIndustry practice

A key judgement: of these 26 items, technical signals are the smallest part. The bulk of the system is cross-respondent aggregation, eligibility gates, institution-level statistics and deliberately unpublished rules. That is why “making one response look genuine” was never the point — the point is that the whole set of respondents is real and independent.

Related cases

  • KAIST (South Korea, 2025) — One department offered overseas academics $100 to take part in a QS-related survey; QS excluded KAIST from the entire 2026 rankings cycle, and the linked responses were never counted. QS release summary ↗
  • University College Cork (Ireland, 2013) — The president asked staff each to recruit three overseas academics to register for the QS survey; QS did not penalise UCC but changed its rules and discarded such institutional sign-ups. Inside Higher Ed ↗
  • Op4G / Slice MR (USA) — Fabricated survey data worth over $10 million using paid fake respondents, VPNs and screener coaching; multiple guilty pleas — cross-industry reference. US DOJ ↗
2

学者身份是否可信Whether an Academic Is Verifiable

三处公开记录交叉核对:ORCID · Scopus · 机构目录Three public records: ORCID · Scopus · Institutional directory ORCID 现职隶属Current ORCID affiliationScopus 发文档案Scopus publication record机构域名邮箱Institutional email domain检查点 8 项8 checkpoints ▾
☑ 标准检查清单 · 8 项标准服务范围内的固定动作
  • ☑ORCID 档案存在且隶属为现职
  • ☑Scopus 发文档案可查且持续有发文
  • ☑机构邮箱域名与官网一致且在用
  • ☑自报隶属 × ORCID × Scopus 三方一致
  • ☑自报熟悉学科与实际发文领域对齐
  • ☑与被评学校无互惠、内部人等关系异常
  • ☑作答环境所在国与自报隶属国一致
  • ☑提名组合无只评一校等异常模式

QS 只能用可机读、可外部核验的数据——核不了的信号,对它等于不存在。以下 8 项就是站在 QS 的位置、用它手边的数据源能建起来的学者核验标准。

核验项怎么对依据
ORCID 档案存在、非空,且隶属与自报一致QS公开
Scopus 发文档案存在发文记录、跨多个年份、近期仍在发文;零发文或新建档案会被标记合理推断
邮箱域名等于机构官方域名;个人或通用邮箱降权或排除QS公开
隶属三方一致自报 × ORCID × Scopus(或机构官网目录)互相吻合行业惯例
熟悉度与发文领域自报熟悉的学科,与本人实际发文领域对得上QS公开
与被评学校的关系合著距离、曾任职、校友身份等关系图谱,可用于识别互惠与内部人(学术界已公开指出互惠偏见问题,QS 未明文承认在用)合理推断
作答环境一致IP 所在国、时区与自报隶属国一致QS公开
作答行为完成时长有下限;提名组合异常(如只评一所、只评关联校群)会被标记行业惯例

对学校的含义:这 8 项全是提交名单前逐人可自查的——ORCID 隶属是否现行、Scopus 是否查得到、机构邮箱是否在用、隶属在官网目录是否可查。信息过期或对不上,是资格层最常见的真实受访者被排除原因,与任何「技巧」无关。

相关案例

  • 雨果奖(2024) — 以假名规律(姓氏按字母序微调、名字为连续数字的译名)识别出 377 张赞助购买的假票并作废,受益入围者本人未被取消资格——跨行业参照。 雨果奖官方报告 ↗
  • Sunday Riley(美国) — 员工以假账号冒充普通用户刷好评、每人建三个不同身份账号,被 FTC 下禁令——跨行业参照。 FTC ↗
  • QS 付费面板试用(2012) — QS 曾通过付费问卷面板试收 400 份学术回复,此后停用现金激励渠道;这是 QS 自身供应链问题,无大学受罚。 Inside Higher Ed ↗
☑ Standard checklist · 8 itemsFixed actions within standard scope
  • ☑ORCID profile exists with a current affiliation
  • ☑Scopus publication record findable and still active
  • ☑Institutional email domain matches the official site and is active
  • ☑Self-report × ORCID × Scopus affiliation all agree
  • ☑Declared subject familiarity matches actual publication areas
  • ☑No reciprocity or insider relationship anomalies with rated institutions
  • ☑Response-environment country matches the declared affiliation country
  • ☑No anomalous nomination patterns (e.g. rating a single institution only)

QS can only use data that is machine-readable and externally verifiable — a signal it cannot check does not exist for it. These eight checks are what QS can build for academics from the data sources at its disposal.

CheckWhat is comparedBasis
ORCID profileExists, non-empty, affiliation matches the self-reportQS-published
Scopus recordPublications exist, span multiple years, still recent; zero-publication or newly created profiles get flaggedInference
Email domainMatches the institution's official domain; personal or generic email is down-weighted or excludedQS-published
Three-way affiliation matchSelf-report × ORCID × Scopus (or the institutional directory) all agreeIndustry practice
Familiarity vs. publication fieldDeclared subject familiarity matches the person's actual publication areasQS-published
Relationship to rated institutionsCo-authorship distance, past employment, alumni status — a relationship graph that can reveal reciprocity and insiders (reciprocity bias is publicly documented in academia; QS has not confirmed using this)Inference
Response-environment consistencyIP country and time zone match the declared affiliation countryQS-published
Response behaviourA floor on completion time; anomalous nomination patterns (rating only one institution, or only a cluster of related ones) get flaggedIndustry practice

What this means for a university: all eight can be checked person by person before a list is submitted — ORCID affiliation current, Scopus record findable, institutional email active, affiliation visible in the official directory. Outdated or mismatched information is the most common reason genuine respondents are excluded at the eligibility layer; it has nothing to do with technique.

Related cases

  • Hugo Awards (2024) — Runs of fake names (surnames altered in alphabetical order, names translating consecutive numbers) exposed 377 sponsor-bought ballots, which were disqualified; the beneficiary finalist was not — cross-industry reference. Hugo administrators' report ↗
  • Sunday Riley (USA) — Staff posted reviews through fake accounts — three identities per person — posing as ordinary customers; the FTC issued a ban — cross-industry reference. FTC ↗
  • QS incentivised panel trial (2012) — QS once trialled sourcing 400 academic responses through a paid survey panel and has not used cash-reward channels since; a QS supply-chain issue — no university was penalised. Inside Higher Ed ↗
3

雇主身份是否可信Whether an Employer Is Verifiable

资格先行 · 真实性核验Eligibility first · Authenticity second 官网与域名年限Website & domain ageLinkedIn 职位核对LinkedIn role match同公司投票密度Voter density per company检查点 7 项7 checkpoints ▾
☑ 标准检查清单 · 7 项标准服务范围内的固定动作
  • ☑公司官网存活且域名有年头
  • ☑LinkedIn 公司页存在且员工数与自报规模对齐
  • ☑个人 LinkedIn 职位与自报一致且账号非新建
  • ☑邮箱域名等于公司官网域名
  • ☑公司注册记录存续可查
  • ☑同公司投票密度无异常集中
  • ☑职位实质参与招聘决策且非教育机构人员
核验项怎么对依据
公司真实存在官网存活;域名注册年限——新注册域名高风险行业惯例
LinkedIn 公司页存在且有年头(业内说法:至少成立 2–3 年,年限阈值未证实合理推断);页面员工数与自报组织规模对得上QS公开
个人 LinkedIn职位与自报一致、账号非新建、连接数异常低会被标记行业惯例
邮箱域名等于公司官网域名;批量同时注册的新域名集群,是壳公司造假的典型形态行业惯例
公司注册记录注册年限与存续状态与各国公司注册库交叉核对行业惯例
同公司投票密度一家公司短期内出现多名雇主受访者,密度异常即标记合理推断
职位实质职位须与「参与招聘决策」匹配;教育机构人员直接无资格QS公开

一个容易忽略的连带风险:来源独立性是按批算的。若一批雇主回答的来源无法核实为完全独立,当年全部相关回答都可被取消资格——所以雇主名单的质量,从来不是某一个人的事。QS公开

相关案例

雇主声誉调查暂无公开点名的处罚案例,以下为机制相近的跨行业参照。

  • Devumi(美国) — 出售机器人假粉丝、播放量等虚假声誉指标,被 FTC 禁止并判 250 万美元——跨行业参照。 FTC 裁决 ↗
  • Roomster(美国) — 购买数万条假好评冒充真实用户评价,被 FTC 与六州永久禁止购买或激励评论——跨行业参照。 伊利诺伊州总检察长 ↗
☑ Standard checklist · 7 itemsFixed actions within standard scope
  • ☑Company website live on an established domain
  • ☑LinkedIn company page exists; headcount matches declared organisation size
  • ☑Personal LinkedIn role matches the self-report; account not newly created
  • ☑Email domain matches the company website domain
  • ☑Company registration active and verifiable
  • ☑No abnormal voter density from a single company
  • ☑Role genuinely takes part in hiring decisions; not education-sector staff
CheckWhat is comparedBasis
Company genuinely existsWebsite live; domain age — a newly registered domain is high-riskIndustry practice
LinkedIn company pageExists and established (industry accounts suggest at least 2–3 years; the threshold is unconfirmedInference); page headcount matches the declared organisation sizeQS-published
Personal LinkedInRole matches the self-report; account not newly created; abnormally low connection counts get flaggedIndustry practice
Email domainMatches the company website domain; clusters of domains registered in bulk at the same time are the classic shell-company patternIndustry practice
Company registrationIncorporation age and active status cross-checked against national registriesIndustry practice
Voter density per companyMany employer respondents from one company in a short window is a density anomalyInference
Substance of the roleThe title must match “takes part in hiring decisions”; education-sector staff are directly ineligibleQS-published

An easily missed knock-on risk: independence of origin is judged per batch. If a set of employer responses cannot be verified as fully independent, all related responses for that year may be disqualified — the quality of an employer list is never about any single person.QS-published

Related cases

No publicly named employer-survey fraud case has been verified; the following are mechanism-adjacent cross-industry references.

  • Devumi (USA) — Sold bot-based fake followers, views and likes — fabricated reputation indicators; the FTC banned the practice and imposed a $2.5m judgment — cross-industry reference. FTC order ↗
  • Roomster (USA) — Bought tens of thousands of fake reviews posing as genuine users; the FTC and six states imposed a permanent ban on buying or incentivising reviews — cross-industry reference. Illinois Attorney General ↗
4

作答环境是否一致Whether the Response Environment Is Consistent

IP · 位置 · 设备配置一致性IP · Location · Device consistency IP × 时区 × 语言一致IP × Time zone × Language156 项检测全集156 checks mapped公开自检工具Public self-check tool检查点 8 项8 checkpoints ▾
☑ 标准检查清单 · 8 项标准服务范围内的固定动作
  • ☑出口 IP 归属地与自报国家一致
  • ☑IP 类型为住宅或机构网络,非机房 / VPN 出口
  • ☑时区 × 语言 × IP 国家三方一致
  • ☑WebRTC / DNS 出口与 HTTP 出口一致
  • ☑设备指纹无自动化与无头浏览器痕迹
  • ☑浏览器语言与系统区域设置一致
  • ☑同一设备或网络无多人集中作答
  • ☑使用日常设备与网络,已关闭 VPN

以下为公开检测工具(who.1ms.dog、CreepJS、BrowserLeaks、BrowserScan、Pixelscan、Whoer、EFF、AmIUnique 等)检测项的并集口径——代表「一份回答的作答环境最多能被读出多少信息」。QS 明文确认采集的是其中 IP、位置、设备配置三类;其余是环境自洽的完整参照系。

A · 服务器可见信息出口 IP(IPv4/IPv6 分列)、IP 归属地(国家/州/市/经纬度)、ASN/ISP/组织、IP 用途分类(住宅/机房/移动/托管)、反向解析主机名、最近边缘节点、IP 所在时区、Tor 出口节点比对、Cloudflare WARP 状态、时钟偏差、服务器日志视角

B · 协议指纹TLS 版本与密码套件、完整 ClientHello、JA3/JA3n 哈希、JA4 系列、ECH 支持、后量子密钥交换降级、SNI 状态、混合内容处理、TCP SYN 指纹(系统家族/TTL/MTU)、HTTP/2 指纹、QUIC/HTTP/3 指纹、HTTP 协议版本

C · DNS / WebRTC / 多出口一致性DNS 解析器地址与归属国、DNS 多查询去重、WebRTC 全部候选地址、WebRTC 公网地址与 HTTP 出口比对、局域网地址、IPv6 推导硬件地址、多站点出口对照、IPv4 与 IPv6 分别测出口、种子下载出口、备用端口出口、延迟差推断中间代理、授权定位与 IP 定位差

D · 请求头User-Agent 原文、Accept/语言/编码、浏览器高低熵信息全套、DNT/全局隐私控制、来源页、代理转发头、AI 代理签名、全部自定义头、请求头与页面读数互校

E · 语言与地区浏览器语言列表、国际化区域设置、系统时区多路互校、语言数据规模、键盘布局、朗读语音列表、语言五路对照(请求头/语言列表/主语言/语音/区域设置)、系统语言包推断

F · 设备指纹面Canvas 哈希、文字平滑、WebGL 参数与渲染图、图形后端一致性、音频指纹、字体列表与哈希、数学函数指纹、DOM 尺寸测量、SVG 尺寸、文字测量框、Emoji 渲染、脚本引擎识别与真实版本、WebAssembly 特性、CSS 特性集、引擎接口计数、计时精度、调用栈深度、媒体格式支持集、综合指纹哈希、模糊/影子指纹、跨浏览器硬件编号、缓存编号、多点持久化标识、防指纹随机化检测

G · 设备信息系统/浏览器/CPU 架构、设备型号推断、屏幕与窗口参数、刷新率、多显示器、GPU 厂商与渲染器、虚拟机判定、WebGPU 能力、WebGL 能力、CPU 线程与内存、触控与指针、电池状态、网络连接类型、媒体设备计数、音频硬件、生物识别能力、蓝牙与外设接口、游戏手柄与 XR、音视频编解码与 DRM、存储配额

H · 浏览器设置与痕迹深浅色与动效偏好、无障碍设置、默认字号与强调色、Cookie 与存储可用性、隐私沙盒接口、权限状态全项、敏感接口暴露计数、PDF 阅读器与插件、内置 AI 可用性、广告拦截检测、追踪拦截实测、内容过滤器识别、扩展痕迹、扩展文件扫描、历史与存储留痕、跨站追踪实测、照片元数据读取、运动传感器

I · 自动化痕迹自动化标记三处检测(主页面/内嵌页/后台线程)、无头浏览器 UA 标记、浏览器对象完整性、插件列表异常、语言列表为空、权限状态互校、Selenium 痕迹、自动化框架点名(Phantom/Playwright/Puppeteer 等)、调试协议检测、调试鼠标泄露、无头相似度百分比、Electron 痕迹、指针行为(无轨迹直达点击)、行为速度(超人速打字填表)、定时分辨率异常

J · 一致性裁决原生函数篡改检查、新鲜页面原生引用对照、后台线程与主页面对照、重复执行一致性、不可能值收集、系统谎言推断(声称的系统 vs 渲染推断的真实系统)、UA 与显卡互校、UA 与引擎版本互校、浏览器信息多路互校、时区/语言/DNS/WebRTC 与 IP 国家互校、传输层与 UA 互校、指纹稳定性裁决与防护方识别、多家评分形态(信任分/真实性百分比/一致性二值/等级)、每条裁决附修复建议

K · 人群对照与回访「在 N 人中是否独特」总判定、每项属性同值占比、每项属性信息熵、与人群的融合度、匿名计数、本地回访识别、跨会话访客 ID 与访问历史

L · IP 信誉与代理判定机房/住宅/移动分类、IP 黑名单查询、代理头与透明代理判定、公开 VPN 与中继判定、IP 历史计数、欺诈评分、无痕模式检测

配套自检工具:覆盖到哪、边界在哪

  • 已覆盖:156 项清单全列全采(浏览器端);自建探针第一期已上线,可给出出口 IP、归属地、网络类型、IP 时区、连接协议等真实读数;结论按三档风险呈现,不编造总分;全程零回传、零统计。
  • 推进中:探针第二期(协议指纹、传输层指纹、DNS 多点连查等约 35 项)随自有服务器一并完成。
  • 主动不做:人群实时对照与 IP 信誉评分——两者都要接外部数据库,与零回传原则冲突,也不暗示 QS 在用。
  • 原理上做不到:26 项标准中有 18 项在 QS 侧计算(跨人判重、来源核实、机构层异常、机器学习验证、资格核实、5 年回溯),任何自检工具都够不到。自检只能降低「真实受访者因自身环境矛盾被误伤」的概率——不承诺符合 QS 要求,更不保证任何一份回答被计入。

配套工具(常年公开、不定向分发):whoami 环境自检工具

相关案例

  • Op4G / Slice MR(美国) — 辅导假受访者用 VPN 掩盖真实 IP 与位置、控制作答时长,近十年绕过纯技术检查——跨行业参照。 美国司法部 ↗
  • Sunday Riley(美国) — 平台删除员工评论后,改用 VPN 隐藏 IP 与位置继续刷评——IP 掩盖本身成为违规信号,跨行业参照。 FTC ↗
☑ Standard checklist · 8 itemsFixed actions within standard scope
  • ☑Exit-IP location matches the declared country
  • ☑IP class is residential or institutional — not a datacentre / VPN exit
  • ☑Time zone × language × IP country all agree
  • ☑WebRTC / DNS exits match the HTTP exit
  • ☑No automation or headless-browser traces in the device fingerprint
  • ☑Browser language matches the system locale
  • ☑No group completion from a single device or network
  • ☑Usual device and network used, VPN switched off

The union of checks performed by public test suites (who.1ms.dog, CreepJS, BrowserLeaks, BrowserScan, Pixelscan, Whoer, EFF, AmIUnique and others) — i.e. how much information a response environment can reveal at most. QS explicitly confirms collecting three of these categories (IP, location, device configuration); the rest form the complete reference frame for environmental self-consistency.

A · Server-visible informationExit IP (IPv4/IPv6 separately), IP geolocation (country/state/city/coordinates), ASN/ISP/organisation, IP usage class (residential/datacentre/mobile/hosting), reverse-DNS hostname, nearest edge node, IP time zone, Tor exit-node matching, Cloudflare WARP status, clock skew, the server-log view

B · Protocol fingerprintsTLS version and cipher suites, full ClientHello, JA3/JA3n hashes, the JA4 family, ECH support, post-quantum downgrade, SNI status, mixed-content handling, TCP SYN fingerprint (OS family/TTL/MTU), HTTP/2 fingerprint, QUIC/HTTP/3 fingerprint, HTTP protocol version

C · DNS / WebRTC / multi-exit consistencyDNS resolver address and country, multi-query DNS de-duplication, all WebRTC candidates, WebRTC public address vs. HTTP exit comparison, local-network addresses, IPv6-derived hardware address, multi-site exit comparison, IPv4 and IPv6 exits measured separately, torrent exit, alternate-port exits, latency-gap proxy inference, authorised geolocation vs. IP location gap

D · Request headersRaw User-Agent, Accept/language/encoding, full low- and high-entropy browser hints, DNT/Global Privacy Control, referrer, proxy forwarding headers, AI-agent signatures, all custom headers, headers cross-checked against page-side readings

E · Language and regionBrowser language list, internationalisation locale, multi-source time-zone cross-checks, locale data size, keyboard layout, speech voices, five-way language comparison (headers/language list/primary language/voices/locale), system language-pack inference

F · Fingerprint surfaceCanvas hash, text smoothing, WebGL parameters and render readback, graphics-backend consistency, audio fingerprint, font list and hash, math-function fingerprint, DOM geometry, SVG geometry, text-measurement boxes, emoji rendering, script-engine identification and true version, WebAssembly features, CSS feature set, engine interface counts, timer resolution, stack depth, media format support, composite fingerprint hash, fuzzy/shadow fingerprints, cross-browser hardware ID, cache-based IDs, multi-point persistent IDs, anti-fingerprint randomisation detection

G · Device informationOS/browser/CPU architecture, device-model inference, screen and window parameters, refresh rate, multiple displays, GPU vendor and renderer, virtual-machine detection, WebGPU capability, WebGL capability, CPU threads and memory, touch and pointer, battery status, connection type, media-device counts, audio hardware, biometric capability, Bluetooth and peripheral interfaces, gamepads and XR, media codecs and DRM, storage quota

H · Browser settings and tracesColour-scheme and motion preferences, accessibility settings, default font size and accent colour, cookie and storage availability, Privacy Sandbox interfaces, full permission states, sensitive-interface exposure count, PDF reader and plugins, built-in AI availability, ad-blocker detection, tracker-blocking tests, content-filter identification, extension traces, extension file scanning, history and storage residue, cross-site tracking tests, photo metadata reading, motion sensors

I · Automation tracesAutomation flags checked in three contexts (main page/embedded page/background worker), headless-browser UA markers, browser-object integrity, plugin-list anomalies, empty language list, permission cross-checks, Selenium traces, automation frameworks named (Phantom/Playwright/Puppeteer…), debugging-protocol detection, debugging mouse leakage, headless-similarity percentages, Electron traces, pointer behaviour (pathless direct clicks), behaviour speed (superhuman typing/form-filling), timer-resolution anomalies

J · Consistency verdictsNative-function tampering checks, fresh-frame native-reference comparison, worker vs. main-page comparison, repeated-execution consistency, impossible-value collection, OS lie detection (claimed OS vs. rendering-inferred OS), UA vs. GPU, UA vs. engine version, multi-source browser-hint cross-checks, time zone/language/DNS/WebRTC vs. IP country, transport layer vs. UA, fingerprint-stability verdicts and protection identification, several scoring shapes (trust score/authenticity percentage/binary consistency/letter grade), a fix suggestion attached to every verdict

K · Crowd comparison and revisits“Unique among N” overall verdict, same-value ratio per attribute, information entropy per attribute, crowd-blending, anonymous counting, local revisit recognition, cross-session visitor ID and visit history

L · IP reputation and proxy verdictsDatacentre/residential/mobile classification, IP blocklist lookups, proxy headers and transparent-proxy verdicts, public VPN and relay verdicts, IP history counts, fraud scores, private-mode detection

The companion self-check tool: coverage and boundary

  • Covered: all 156 checks listed and collected in the browser; a first phase of self-built probes is live, returning real readings for exit IP, geolocation, network type, IP time zone and connection protocol; results are presented in three risk tiers with no invented composite score; zero data return, zero analytics throughout.
  • In progress: probe phase two (protocol fingerprints, transport-layer fingerprints, multi-point DNS checks — roughly 35 items), delivered with a dedicated server build.
  • Deliberately not built: live crowd comparison and IP reputation scoring — both require external databases, conflict with the zero-return principle, and must not imply QS uses them.
  • Impossible in principle: 18 of the 26 standards are computed on QS's side (cross-respondent de-duplication, provenance verification, institution-level anomalies, ML validation, eligibility verification, 5-year retrospection) and no self-check tool can reach them. A self-check can only lower the chance that a genuine respondent is caught out by contradictions in their own environment — it promises neither QS compliance nor that any response will be counted.

Companion tool (public year-round, never distributed through any institution): the whoami environment self-check

Related cases

  • Op4G / Slice MR (USA) — Fake respondents were coached to mask real IP and location with VPNs and to control dwell time, defeating purely technical checks for nearly a decade — cross-industry reference. US DOJ ↗
  • Sunday Riley (USA) — After the platform removed employee reviews, the company used a VPN to hide IP and location and kept posting — the masking itself became an evasion signal; cross-industry reference. FTC ↗
5

参与规范与常见误区Participation Standards & Common Misconceptions

中性提醒 · 独立作答Neutral reminder · Independent response 官方模板提醒Official-template reminder禁止请托与激励No solicitation or incentives激励判例:整校排除Incentive precedent: full exclusion检查点 7 项7 checkpoints ▾
☑ 标准检查清单 · 7 项标准服务范围内的固定动作
  • ☑使用 QS 官方模板且一对一发送
  • ☑发送前已取得对方同意
  • ☑提醒文案中性,无请托措辞
  • ☑无现金、礼品等任何形式的激励
  • ☑无填写指南、提名建议等指导作答
  • ☑无代填,无截图回报要求
  • ☑无同一时间、同一网络的集中填写组织

良好实践(可以做)

  • 用 QS 官方模板邮件、一对一发送,事先取得对方同意QS公开
  • 名单走 QS Hub 提交,或用 QS 官方公开注册链接,二选一QS公开
  • 提醒保持中性:一句提醒 + 官方链接 + 截止时间QS公开
  • 事前给所有人一份通用环境说明:用平常的设备和网络填写、关掉 VPN、不用酒店或网吧公共网络合理推断
  • 提醒名单里的人保持学术与职业身份信息最新:ORCID 隶属现行、机构邮箱在用、LinkedIn 职位与现职一致合理推断

红线(不能做)

  • 不写「请投票支持本校」,不附任何填写指南或提名建议——指导作答(coaching)是 QS 明文禁止项QS公开
  • 不给任何激励:现金、礼品、好处都不行QS公开
  • 不代填、不要求截图回报、不组织同一时间、同一网络集中填写QS公开
  • 任何工具与材料都不得与某个机构的触达链关联——来源可追溯,风险会连带整批回答QS公开

常见误区

  • 「熟悉度过滤可以绕开」——不行,也不应尝试:声明之外的评价会被过滤或降权,这是问卷设计本身。QS公开
  • 「名单里有合作关系才算数」——科研合作是最硬的真实关系证据,但会议伙伴、外部评审、聘用过毕业生的雇主同样是 QS 鼓励的真实关系,不必有合著。
  • 「某个人的票没算上,一定是关系不够」——更常见的原因在资格层:类别放错、身份无法核实、信息过期(见下方案例)。
判例 · 激励的代价:KAIST 曾因一封带 100 美元激励的邮件,整校被排除出一个排名周期。激励不需要真的发出多少——一封邮件就够了。
案例 · 类别与资格(据转述,细节待核实):某合作教育机构与一所大学有正式合作,其人员的评价最终未被计入。更可能的机制不在「关系」,而在资格:其一,相关人员属关联方,可能被视为「自己人」,撞上自投排除与来源独立性;其二,若列入雇主名单,教育机构人员明文无雇主资格;其三,若列入学者名单,实务背景、无公开科研档案,学者身份无法核实。QS 核实的是投票人本人的资格,不是投票人与学校的合作关系。

相关案例

  • 科克大学 UCC(爱尔兰,2013) — 动员注册未被认定违反当时规则,但 QS 就此禁止机构自行推广注册——善意动员与违规拉票的界线由此划定。 Inside Higher Ed ↗
  • 都柏林圣三一(爱尔兰,2016) — 信件附 QS 注册链接被 QS 认定违规并受谴责,受影响回复将被排除;更进一步的停榜处置仅称「将考虑」,未见最终证实报道。 Inside Higher Ed ↗
  • 贝勒大学(美国,2008) — 付费让已录取新生重考 SAT 以抬高可报分数,曝光后自行叫停;属激励改变测量行为、非数据造假,无 U.S. News 处罚。 Inside Higher Ed ↗
☑ Standard checklist · 7 itemsFixed actions within standard scope
  • ☑QS official template used, sent one-to-one
  • ☑Prior consent obtained before sending
  • ☑Reminder wording neutral, with no solicitation language
  • ☑No incentives in any form — cash, gifts or favours
  • ☑No completion guidance or nomination suggestions (no coaching)
  • ☑No proxy completion and no screenshot reporting
  • ☑No organised same-time, same-network group completion

Good practice

  • Use QS's official template email, sent one-to-one, with prior consentQS-published
  • Submit lists through the QS Hub, or use QS's official public sign-up link — one route or the otherQS-published
  • Keep the reminder neutral: one line, the official link, the deadlineQS-published
  • Give everyone the same general environment note in advance: use your usual device and network, turn the VPN off, avoid hotel or internet-café public networksInference
  • Remind listed contacts to keep academic and professional identity information current: ORCID affiliation up to date, institutional email active, LinkedIn role matching the current postInference

Red lines

  • No “please support this university”, and no completion guidance or nomination suggestions of any kind — coaching is expressly prohibited by QSQS-published
  • No incentives of any form: cash, gifts or favoursQS-published
  • No proxy completion, no screenshot reporting, no organised same-time, same-network group completionQS-published
  • No tool or material may be traceably linked to any institution's outreach chain — provenance is recorded, and the risk attaches to the whole batch of responsesQS-published

Common misconceptions

  • “The familiarity filter can be worked around.” It cannot, and should not be attempted: ratings outside declared familiarity are filtered or down-weighted by the survey's own design.QS-published
  • “Only formal collaborators count.” Research collaboration is the hardest evidence of a genuine relationship, but conference partners, external examiners and employers who have hired the university's graduates are equally relationships QS encourages — no co-authorship required.
  • “A vote didn't count because the relationship was too weak.” The more common causes sit at the eligibility layer: wrong category, unverifiable identity, outdated information (see the case below).
Precedent · the price of an incentive: KAIST was once excluded from an entire rankings cycle over a single email offering a $100 incentive. The incentive does not need to reach many people — one email is enough.
Case · category and eligibility (as relayed; details unverified): a partner education institution in formal partnership with a university saw its people's ratings ultimately not counted. The likely mechanism is not “relationship” but eligibility: first, the individuals concerned were affiliated parties, so they may have counted as insiders — hitting self-nomination exclusion and independence of origin; second, if listed as employers, education-sector staff are expressly ineligible as employers; third, if listed as academics, practitioner backgrounds without a public research record leave the academic identity unverifiable. QS verifies the respondent's own eligibility, not the respondent's partnership with the university.

Related cases

  • University College Cork (Ireland, 2013) — The recruitment drive was held not to infringe the then-policy, but QS responded by banning independent institutional promotion — drawing the line between good-faith mobilisation and canvassing. Inside Higher Ed ↗
  • Trinity College Dublin (Ireland, 2016) — Letters carrying the QS sign-up link were found by QS to breach its guidelines and were censured; influenced respondents were to be excluded, while further sanctions were only stated as “to be considered” — no final suspension is confirmed in reporting. Inside Higher Ed ↗
  • Baylor University (USA, 2008) — Paid already-admitted freshmen to retake the SAT to lift its reportable average, then abandoned the scheme within days; gaming that changes the measured behaviour rather than falsifying it — no U.S. News sanction. Inside Higher Ed ↗
6

名单与数据的日常治理List & Data Governance, Day to Day

三筛 · 逐人自查Three screens · Person-by-person check 非内部人员Not insiders类别无错位Correct category身份可核实Individually verifiable检查点 8 项8 checkpoints ▾
☑ 标准检查清单 · 8 项标准服务范围内的固定动作
  • ☑本校及关联方人员已筛除
  • ☑学者 / 雇主分类无错位
  • ☑教育机构人员未列入雇主名单
  • ☑逐人身份可核实:邮箱在用、公开档案可查
  • ☑名单总数未超每项调查 400 人上限
  • ☑名单在两年有效期内且已更新
  • ☑与学校有可留痕的真实关系
  • ☑重复与过期联系人已清理

提交前三筛

  • 不是自己人:本校人员、关联机构及关联方人员不提交——会撞自投排除与来源独立性QS公开
  • 分类别不错:学者进学者名单、雇主进雇主名单;教育机构人员不进雇主名单(明文无资格)QS公开
  • 本人能被核实:机构邮箱在用、公开档案查得到、隶属是现职QS公开

量化自查清单(逐人可核验,不是凭感觉)

  • 学者:ORCID 有且隶属现行;Scopus 查得到发文档案;机构邮箱在用;隶属在机构官网目录可查;非本校及关联机构人员。
  • 雇主:公司官网存活且域名有年头;LinkedIn 公司页存在且非新建(业内说法至少 2–3 年,未证实合理推断);本人职位在 LinkedIn 或官网对得上;邮箱域名与官网一致;非教育机构人员。
  • 关系:与学校有可留痕的真实关系——合著、会议、评审、聘用毕业生,任一即可,不必全是合著。

名单的硬规则

  • 每校每项调查提名封顶 400 人,公开注册同样封顶QS公开
  • 名单 两年有效,到期须更新——这也是身份信息要保持现行的原因QS公开
  • 调查由 QS 从官方邮箱直接发送(1 封邀请 + 2 封提醒),受访者独立完成;学校能做的只有提交名单与中性提醒QS公开

相关案例

以下均为机构数据造假案例,机制与声誉投票操纵不同,列此对照名单与数据治理失守的后果。

  • 哥伦比亚大学(美国,2022) — 提交的班级规模、学位等数据不实,被 U.S. News 取消当年排名——机构数据造假。 Inside Higher Ed ↗
  • 伊利诺伊大学法学院(美国,2012) — 虚报 LSAT 与录取数据,被 ABA 公开谴责并罚款 25 万美元——机构数据造假。 Inside Higher Ed ↗
  • 天普大学 Fox 商学院(美国,2021) — 院长因多年提交虚假排名数据被判电信欺诈罪、入狱 14 个月——机构数据造假的刑事后果。 Inside Higher Ed ↗
☑ Standard checklist · 8 itemsFixed actions within standard scope
  • ☑Own and affiliated-party staff screened out
  • ☑Academic / employer categories correctly assigned
  • ☑No education-sector staff on the employer list
  • ☑Each person individually verifiable: active email, findable public profile
  • ☑List within the 400-per-survey cap
  • ☑List inside its two-year validity and refreshed
  • ☑A genuine, traceable relationship to the university on record
  • ☑Duplicate and outdated contacts removed

The three screens before submission

  • Not insiders: the university's own staff, affiliated institutions and affiliated-party personnel are not submitted — they collide with self-nomination exclusion and independence of originQS-published
  • Right category: academics on the academic list, employers on the employer list; education-sector staff never on the employer list (expressly ineligible)QS-published
  • Individually verifiable: institutional email active, public profile findable, affiliation is the current postQS-published

The quantifiable checklist (verifiable per person, not by feel)

  • Academics: ORCID exists with current affiliation; a Scopus publication record is findable; institutional email active; affiliation visible in the institution's official directory; not staff of the university or its affiliates.
  • Employers: company website live on an established domain; LinkedIn company page exists and is not new (industry accounts suggest at least 2–3 years, unconfirmedInference); the person's role matches on LinkedIn or the website; email domain matches the website; not education-sector staff.
  • Relationship: a genuine, traceable link to the university — co-authorship, conferences, examining, hiring its graduates; any one suffices, it need not be co-authorship.

The hard rules of the list

  • Nominations capped at 400 per institution per survey; public sign-ups are capped the same wayQS-published
  • Lists remain valid for two years and must then be refreshed — one more reason identity information has to stay currentQS-published
  • Surveys are sent by QS directly from its official address (1 invitation + 2 reminders) and completed independently; all a university can do is submit the list and send a neutral reminderQS-published

Related cases

All of the following are institutional data-misreporting cases — a different mechanism from reputation-survey manipulation — included to show what failed list and data governance costs.

  • Columbia University (USA, 2022) — Inaccurate class-size and degree data led U.S. News to unrank Columbia for that cycle — institutional data misreporting. Inside Higher Ed ↗
  • University of Illinois College of Law (USA, 2012) — Intentionally inaccurate LSAT and admissions data brought an ABA public censure and a $250,000 sanction — institutional data misreporting. Inside Higher Ed ↗
  • Temple University, Fox School of Business (USA, 2021) — The dean was convicted of wire fraud and sentenced to 14 months in prison over years of false rankings data — the criminal end of institutional data misreporting. Inside Higher Ed ↗
7

回答内容与群体模式Response Content & Group Patterns

相关性 · 一致性 · 群体模式Relevance · Consistency · Collective patterns 提名与专长对齐Nominations match expertise答案前后一致Internally consistent answers集中涌入与高度雷同Bursts & near-identical sets检查点 8 项8 checkpoints ▾
☑ 标准检查清单 · 8 项标准服务范围内的固定动作
  • ☑提名与受访者申报专长学科对齐
  • ☑提名地域与受访者熟悉地域对齐
  • ☑自投本校已排除
  • ☑回答与身份档案无前后矛盾
  • ☑无直线作答、模式化作答痕迹
  • ☑提交时间无异常集中涌入
  • ☑提名组合无高度雷同、无只推一校
  • ☑单一信号不作结论,异常先做统计调节复核

前几块查的是「人对不对、环境对不对」;这一块查答案本身:提名与作答者的身份、专长对不对得上,一群人的答案是不是高度雷同、集中涌入、齐刷刷只推同一所学校——这是识别批量灌票(survey farms)最核心的一层。三组证据强度不同,标注分开。

第一组 · 相关性:提名与这个人对不对得上(QS 明文为主)

  • 受访者先申报专长学科与地域熟悉度,申报决定其能提名哪些学校;超出专长 / 地域范围的评价被过滤或降权QS公开
  • 自投排除:学者无法在问卷中选择自己所属机构QS公开
  • 资格定义决定相关性:学者须从事教学或研究,雇主须实际参与招聘决策QS公开
  • 提名学校与本人发文领域 / 从业领域明显无关,是相关性复核的触发点合理推断

第二组 · 一致性:答案与本人、与自身是否自洽(行业标准,QS 未逐项点名)

  • 答案与身份档案矛盾:隶属、职位、领域与公开档案对不上行业惯例
  • 同一受访者前后答案自相矛盾(连贯性评分是调研质检的标准项)行业惯例
  • 直线作答 / 模式化作答:量表题全选同一项、呈机械规律行业惯例
  • 作答时长异常短、开放题无实质内容或与他人雷同行业惯例

第三组 · 群体模式:一群人的答案像不像商量好的

  • 异常检测盯非典型回答模式与某校表现的异常跃升;不合参数的回答被移除,异常机构被深入审查QS公开
  • 一批回答的来源若无法核实为完全独立,当年全部相关回答可被取消资格QS公开
  • 多份回答的答案向量高度相同、开放题文本重复,是行业代码框架明列的剔除理由行业惯例
  • 「判定为针对某校的协同灌票」是多信号的组合推断——QS 未公布任何份数、相似度阈值,任何具体数字都不是 QS 标准合理推断
单一信号不作结论:提名集中本身不等于作弊——学校正常提交联系人名单(QS 允许、每校每项调查上限 400 人)天然会产生集中与时间聚集。QS 公开的处置顺序是先做统计调节、压低异常高提名量的影响,再谈剔除与处罚;把任何单一信号直接读成「刷票」,既不符合 QS 的做法,也冤枉正常名单。

相关案例

  • 雨果奖(2024) — 针对同一入围者的集中投票呈假名规律与赞助购买模式,377 张票被作废——群体模式识别的标杆案例,跨行业参照。 雨果奖官方报告 ↗
  • Roomster(美国) — 数万条 4–5 星假评以量与星级分布淹没真实评价,被监管查处——答案分布 flooding 的参照,跨行业。 伊利诺伊州总检察长 ↗
  • 贝鲁特美国大学(2025) — 称 THE 排名存在「互投集团」等操纵而自行退出;未点名任何学校、THE 亦未处罚——仅作模式警示,非已证实个案。 AUB 官方声明 ↗
☑ Standard checklist · 8 itemsFixed actions within standard scope
  • ☑Nominations align with the respondent's declared subject expertise
  • ☑Nominations align with the respondent's declared regional familiarity
  • ☑Self-nomination of the home institution excluded
  • ☑Answers show no contradiction with the identity profile
  • ☑No straight-lining or patterned-response traces
  • ☑No abnormal burst of submissions in a short window
  • ☑Nomination sets not near-identical; no single-institution-only pushing
  • ☑No conclusion from a single signal; anomalies go to statistical adjustment first

The earlier blocks ask whether the person and the environment check out; this block examines the answers themselves: whether nominations fit the respondent's identity and expertise, and whether a set of responses looks near-identical, burst-submitted and single-mindedly pushing one institution — the core layer for detecting bulk ballot-stuffing (survey farms). The three groups carry different evidence grades, labelled separately.

Group 1 · Relevance: do the nominations fit this person? (mostly QS-published)

  • Respondents first declare subject expertise and regional familiarity, which determines which institutions they may nominate; ratings outside the declared scope are filtered or down-weightedQS-published
  • Self-nomination excluded: academics cannot select their own institutionQS-published
  • Eligibility defines relevance: academics must teach or research; employers must actually take part in hiring decisionsQS-published
  • Nominations plainly unrelated to the respondent's publication or professional field are a trigger for relevance reviewInference

Group 2 · Consistency: do the answers agree with the person and with themselves? (industry standard; QS has not itemised these)

  • Answers contradicting the identity profile — affiliation, role or field at odds with the public recordIndustry practice
  • Self-contradiction across a respondent's own answers (coherence scoring is a standard survey-QC control)Industry practice
  • Straight-lining / patterned responding: the same option down a whole grid, or mechanical patternsIndustry practice
  • Abnormally short completion times; open ends with no substance or duplicated across respondentsIndustry practice

Group 3 · Group patterns: does a set of answers look coordinated?

  • Anomaly detection watches for atypical response patterns and unusual jumps in an institution's performance; out-of-parameter responses are removed and anomalous institutions reviewed in depthQS-published
  • If a set of responses cannot be verified as fully independent in origin, all related responses for that year may be disqualifiedQS-published
  • Near-identical answer vectors across respondents and duplicated open-end text are explicit removal reasons in the industry's shared code frameIndustry practice
  • “Coordinated stuffing aimed at one institution” is a combined, multi-signal inference — QS publishes no counts or similarity thresholds, and any specific number is not a QS standardInference
No conclusion from a single signal: concentrated nominations are not in themselves cheating — a university's normal contact-list submission (permitted by QS, capped at 400 per survey) naturally produces concentration and time clustering. QS's published order of response is statistical adjustment first, dampening the effect of anomalously high nomination volumes, before removal or sanction; reading any single signal as “ballot-stuffing” matches neither QS practice nor fairness to a normal list.

Related cases

  • Hugo Awards (2024) — Concentrated voting for one finalist showed fake-name runs and a sponsor-refund pattern; 377 ballots were disqualified — the benchmark case for collective-pattern detection; cross-industry reference. Hugo administrators' report ↗
  • Roomster (USA) — Tens of thousands of bought 4–5-star reviews flooded the genuine rating distribution and were sanctioned by regulators — a reference for answer-distribution flooding; cross-industry. Illinois Attorney General ↗
  • American University of Beirut (2025) — Withdrew from THE rankings alleging “reciprocal voting syndicates” among unnamed others; no institution was named and THE sanctioned none — a pattern warning only, not a proven case. AUB official statement ↗

工作边界Work Boundaries

范围:名单干净 · 身份现行 · 合规提醒 · 环境一致 Scope: Clean lists · Current identities · Compliant reminders · Consistent environments

  • 不动员、不组织任何人投票It never mobilises or organises anyone's vote
  • 不指导任何人如何作答It never coaches anyone on how to respond
  • 不承诺任何一份回答会被 QS 计入It never promises that any response will be counted by QS
  • 所有判断区分「QS 明文」与「合理推断」,不把推断说成事实Every judgement separates “QS-published” from “inference”, and never presents inference as fact