Topic
Is passwordless login better?
Verifiable question
For consumer web apps using passkeys in 2025:
• Does signup completion change versus email plus password?
• What recovery failures appear after users lose a device?
• Which browsers and operating-system versions were measured?
Decision: whether to run a passkey pilot, not whether passkeys are “good.”
Claim in blog: “The rule affects 40% of contractors.”
Trace:
blog → newspaper → trade association press release → survey PDF
At the PDF, check:
• Who designed and paid for the survey?
• How were contractors sampled?
• What was the exact question?
• Does 40% mean all contractors or 40% of 312 respondents?
• Is there a table or only a press-release sentence?
SIFT THE SOURCE
Scope Does it address my exact question?
Identity Named author, role, contact, publication history?
Funding Sponsor, affiliate links, employer, political interest?
Method Sample, definitions, exclusions, raw data, limitations?
Trace Can I open the cited evidence?
Correction Date, version history, correction policy?
“Experts say” names nobody.A source that shows its limits is stronger, not weaker.
看起来像证据的搜索结果
有条结果写的是 “Top 10 VPNs for Privacy | We tested every leading provider.” 页面自己的披露声明写着，读者通过它的链接下单它就拿佣金；而它所谓的“测试”既没写协议，也没写服务器、日期和测量方式。修法是：这个页面只用来收集产品名字，然后去看独立的协议审计报告和隐私政策，并且在任何还留着的结论旁边，把这层佣金关系记下来。
Record four clocks when they matter:
PUBLISHED 2024-03-12
UPDATED 2025-09-08
DATA COLLECTED 2023-01 through 2023-12
VERSION PostgreSQL 16.4
Question to ask:
“Was this true on the date of the decision and for our version?”
Do not cite “updated yesterday” when the underlying table is from 2018.
失败案例：步骤没错，版本不对
有位开发照着最新的 API 文档做，拿到的是 “400 Bad Request: Unrecognized field 'functions'.” 那个账号被锁在一个更早的 API 版本上，那版要的是另一个字段名。修法：把端点和版本请求头记下来，打开对应版本的存档文档，比对请求结构，一次只改一个字段，然后把能跑通的请求连日期一起存起来。“文档写错了”这个判断太笼统，没法调。
CLAIM: “Complaints doubled in 2025.”
Source A Company report 1,040 → 2,090 tickets
Source B Regulator database 210 → 260 complaints
Source C Support dashboard changed category rules in July
Reconciliation:
The company count doubled, partly after a classification change.
The regulator count rose 24%. These measures are not interchangeable.
Do not write: “Multiple sources confirm complaints doubled.”
CLAIM ID: C-07
Claim: Vendor will support version 3 through 2028-06-30.
Source: “Version 3 Support Policy,” Acme Docs
Published/updated: 2026-02-10
URL: https://example.com/support/v3
Accessed: 2026-07-25
Evidence: “Security fixes continue through June 30, 2028.”
Location: Long-term support, paragraph 2
Archived URL / local PDF hash:
Notes: Applies to Enterprise plan only.
原话要摘够上下文，把限定条件一起留住。“Security fixes continue through 2028” 这句话，如果下一句把这个承诺限定在某个付费档位，单摘出来就会误导人。截图能保住视觉证据，但可检索的文本加一个稳定的 URL，别人复核起来更省事。
AI answer, copied exactly:
“Smith and Rao (2021), Journal of Digital Methods,
doi:10.1234/jdm.2021.0042, found a 37% accuracy increase.”
DOI lookup:
DOI Not Found
10.1234/jdm.2021.0042
Repair:
1. Search the exact title without the DOI.
2. Search both authors plus the claimed phrase.
3. Search Crossref or the publisher by title and year.
4. If no record appears, mark the citation fabricated and delete the claim.
5. Find real studies from a database, then read methods and results.
危险的地方恰好是那些看着靠谱的细节
那个编出来的 DOI 结构完全正规，期刊名听着平平无奇，百分之三十七这个数字显得很精确。这些一样都不是证据。也别回头问同一个模型“你确定吗”，那只会再生成一句话。离开模型，去出版方、文献索引库或者原始文档那边核。
Original request:
HTTP/1.1 404 Not Found
The requested URL /reports/2024-risk-review.pdf was not found.
Recovery ladder:
1. Search: "2024 Risk Review" publisher name
2. Search: site:publisher.example filetype:pdf "Risk Review"
3. Remove /reports/ from the URL and inspect publications
4. Search the exact filename and report number
5. Check an archived capture of the original URL
6. Ask the publisher for the superseding document
7. Record that the original was unavailable and what replaced it
QUESTION
□ Decision, population, comparison, measure, time period
SEARCH
□ Broad vocabulary → exact phrase → source/domain → document
□ Record useful and failed queries
VERIFY
□ Open the primary source
□ Check author, method, funding, incentives, date, and version
□ Find independent evidence for the important claims
□ Open every AI-suggested citation; match claim to passage
PRESERVE
□ Title, author, publisher, URL, date, access date, quote, location
□ Archive unstable pages lawfully
□ Mark paywall, 404, replacement, and unresolved disagreement
STOP
□ Can another person reproduce the path?
□ Would missing evidence change the decision?
□ Is confidence stated honestly?