How we source what we say

What you get

• A source behind every fact

  Tap a citation number to see where a fact comes from: the page, who published it, and what kind of source it is.

• Checked twice, independently

  Two verifiers fetch each source again and confirm the quote is there and supports the fact. If either cannot, it is cut.

• Only what a source can support

  A job posting shows what an employer advertises, not what happens in its interviews. One post is one person’s account.

• No invented numbers

  No employer publishes a pass rate, a round weight or a typical timeline, so you will not find one here.

SourceStrongAI, the company that runs FDE Interview Prep, publishes this research. Below is the method, as short proof.

How a fact gets in

A researcher reads the source itself: a job posting, a careers page, a company blog, a filing, a news article or a candidate’s post. Each claim is written down with a quote copied word for word from the page, its address, its publisher and its source type.

Then two verifiers check it, each on their own. Each one fetches every source again and checks that the quote is on the page and that it supports the claim. A claim survives only if neither verifier finds it unsupported. Where a source supports only part of a claim, the claim is narrowed to what the source says.

Verification shows that a source says what the claim says, not that the source is right. So every claim carries its source type, and a page states it only in words that type allows.

The corpus holds 2,146 verified claims in 54 research files, backed by 3,820 source citations.

Independent audits

After verification, a fresh auditor took a random sample of 40 claims from 26 research areas, weighted toward the riskier source types, and re-checked each one on the assumption that the earlier checks had been too lenient. Result: 38 correct, two overstated, none wrong. Each overstated claim was corrected.

A second audit re-checked a sample of 60 claims from the second research round: 50 correct, 10 overstated, none wrong, and each overstated claim was narrowed. An independent re-derivation from the raw data reproduced every headline number in the posting census and the H-1B analysis.

Then every candidate-report and prompt claim, 474 in all, was checked again under a stricter rule: a claim may say only what the poster wrote, with the poster’s hedges, and nothing about how often something happens. Result: 233 correct, 241 overstated and narrowed, none wrong.

Claims that a verifier could not confirm are dropped, not softened: 12 were dropped in the second round, most of them posts deleted with no archived copy.

What each kind of source can support

A source is evidence for some statements and not others. These are the limits every page follows.

The employer’s own words

Job postings, careers and hiring pages, company blogs, filings and executives’ own posts.

Can support: What the employer states or advertises: titles, requirements, posted pay, stated travel, published hiring steps and rules on AI use.

Never what happens in a real loop, what people are actually paid, or how common something is across employers.

Candidate reports

Individual accounts on Blind, Reddit, Hacker News, personal blogs and GitHub.

Can support: That one person reported something, with when and the role they named. We write “candidates report” only when two or more independent posters say it.

Never a company’s policy, the usual loop, or how often something happens.

Press

Journalism and newsletters.

Can support: That a named outlet reported something, and quotes from named people.

Not data the outlet only relays, unless we say whose data it is.

Prep and salary sites

Interview prep sites, salary aggregators and job-market trackers.

Can support: Only that the site says something. Self-reported pay always carries its sample size.

Never a company’s policy, its rounds, pass rates, or anything presented as usual.

Our own analysis

Counts we made ourselves from public job boards, and one keyword count of GitHub repositories.

Can support: A count within one snapshot, stated with the boards, the title patterns and the unit (see the posting census).

Not anything outside the boards we read, and not growth: a snapshot is one moment.

Independent analysis

Posting datasets, surveys and field guides with a stated method.

Can support: A share or median within that dataset, stated with its window and method.

Not a comparison with a dataset measured another way, total pay, or anything outside the sample.

Government data

US Department of Labor H-1B filing records.

Can support: The offered base wage on certified filings, as a count, median or percentile, with its window (see the H-1B filings).

Never total pay, what anyone is paid now, or a hire.

Our posting census

In one pass, we fetched 103 employer job boards and kept every live posting whose title matches one of these patterns: “forward deployed”, “deployed engineer”, “deployment strategist”, “agent deployment engineer” and “FDE”. We match titles only, never descriptions.

We read 95 boards whole, from Ashby, Greenhouse, Lever and Rippling, so matching on them is exhaustive. We read the other eight through each site’s own keyword search (Adobe, Amazon, EY, Google, Microsoft, NVIDIA, Salesforce and Workday), so those employers may be undercounted.

The census found 806 postings from 92 employers. Its unit is a posting, not a distinct role: an employer often posts one role in several places, and the 806 postings are 588 employer-and-title pairs.

What it leaves out: employers that publish no job feed we can read, including the consultancies Accenture, Deloitte, PwC, KPMG and IBM. We counted Accenture and Deloitte from their own careers sites, and those counts are kept apart from the census. Its counts are not comparable with trackers that use other methods.

So a census figure on this site always reads as postings on the job boards we checked, when we counted.

H-1B salary filings

Before an employer sponsors a worker on an H-1B, H-1B1 or E-3 visa, it files a Labor Condition Application with the US Department of Labor, stating the job title, the worksite and the wage it will pay. The Department publishes these filings as disclosure files.

We read five of those files, which hold seven quarters of decisions. Where a later file already holds an earlier one’s decisions, we read the later file and use the earlier one only as a check.

We keep certified filings whose job title contains one of five patterns (“deployed engineer”, “deployment engineer”, “deployment strategist”, “forward deployed” and “forward-deployed”), and map each worksite to a metro area with the Census Bureau’s delineation file. That gives 607 certified filings, 387 of them with a forward-deployed title.

What a filing records is the offered base wage: the lower bound of the wage the employer stated, put on a yearly basis. It has no field for bonus, equity or commission, so it is never total pay. It covers sponsored workers only. A certified filing is not a hire, and an FDE filed under another title is missed. The broad “deployment engineer” pattern also catches roles that are not forward deployed, so we prefer the forward-deployed titles. We never compare filing wages with posted pay bands.

Where the source links go

Each citation links to the page a reader can check. Where we read a posting through a job board’s data feed, the link goes to the posting itself, or to its Wayback Machine snapshot once it has closed. When a claim covers several postings or a whole board, the link goes to the employer’s job board. Claims from our own analysis link to the census or the H-1B method above.

Found a fact that looks wrong? Tell us, and we will check it against its source. About says who we are.