Method, test results and sources

Last updated November 9, 2026. Engine 1.x. Every rule is public here so you can judge it.

Principles

Only use what the candidate gave you, plus free public lookups (DNS, the domain registry, GitHub's public API).
Every finding comes with evidence and a way to verify it. The output is a checklist, never a decision.
Never use protected traits, and avoid known stand-ins for them. We do not use employment gaps, age, graduation year, photos, accent, nationality, or "AI-written" scores, and we never read a name as a hint about origin, sex or age. Names are only compared with each other, such as the name you typed against the name on the resume.
Prefer signals that are cheap to verify and fair to honest people.

What we check

Finding	Level	Why
Disposable email domain, email domain that does not exist	Verify first	Throwaway or impossible contact details
Domain that cannot receive mail, custom domain registered in the last 180 days	Worth verifying	Fresh or broken domains are common in made-up identities, and sometimes honest
Phone number we cannot read, or in the reserved 555-01XX fiction range	Worth verifying	Not a reachable line. Numbers typed as VoIP in public numbering data are shown as context only, because US VoIP numbers are rarely typed that way
The name at the top of the resume differs from the name you entered	Worth verifying	Borrowed resume or identity; also married names and templates
Resume lists a different email, phone, LinkedIn or GitHub	Worth verifying	Stitched-together identities
GitHub profile missing; very new account behind senior claims; profile handle or display name is a different person	Verify first / Worth verifying	Borrowed or synthetic online presence
Unfilled template text, chatbot replies left in	Verify first	Nobody read the resume before sending it. Using AI to write is fine and not flagged
Text that addresses an automated screener, hidden tiny text in the PDF	Verify first	Deliberate manipulation
More years of a technology than it has existed (Verify first); a tool used in a job that ended before it existed (Worth verifying). We allow one extra year for rounding	Mixed	Fabricated history
Dates that end before they start; two full-time jobs overlapping a year or more	Worth verifying	Fabrication, or honest moonlighting. Part-time, contract and per diem roles are excluded
Two or more employer names so generic we cannot tell which company they are, with no location	Worth verifying	Hard to verify; common in fabricated histories
Job description copied word for word (25% or more of its 7-word phrases)	Worth verifying	Hides what the person actually did
Text that asks for pay into another person's account, equipment shipped to someone else, or offers remote access to a work laptop (Verify first); pay in crypto, camera unavailable, ship-to address in another state (Worth verifying)	Mixed	Published FBI and Microsoft guidance on identity-masking remote workers. We look for request-style wording to avoid matching job duties. It can still misfire, which is why every finding is a question
A UTC or GMT offset in an availability statement that does not fit a US work location	Worth verifying	About where they say they will work, not where they are from. Bare zone abbreviations are ignored, and lines inside past job entries are skipped where we can tell
PDF author field has a different name; 60+ item skills list; many short engagements	Context	Weak on their own

Test results

We tested on made-up resumes written for this purpose: a development set we tuned rules on, and a separate holdout set written independently afterwards that we did not tune on. Most test phone numbers use the reserved 555-01XX fiction range by design, so that one finding is left out of these numbers. Results below are without network lookups, because the made-up custom domains do not exist and would inflate both catches and false alarms.

Development set (40 samples)

Rules were adjusted while looking at this set, so these numbers are optimistic.

Counted as flagged when the report says	Fakes caught	Honest candidates flagged
Verify first	7 of 18	0 of 22
Verify first, or two or more Worth verifying	10 of 18	0 of 22
Any finding shown	15 of 18	0 of 22

Missed fakes: fake_custom_domain; keyword_stuffing; impossible_timeline.

Holdout set, frozen rules (24 samples)

Written independently by a separate process from a prompt that did not describe the tool, and run once with the rules frozen (engine 1.1). This is the honest number.

Counted as flagged when the report says	Fakes caught	Honest candidates flagged
Verify first	1 of 11	0 of 13
Verify first, or two or more Worth verifying	2 of 11	0 of 13
Any finding shown	3 of 11	0 of 13

Missed fakes: overseas operator posing as US-based remote developer (laptop forwarding); fabricated and inflated work history (technology anachronisms); third party (bench sales recruiter) applying on the candidate's behalf; fabricated work history backed by a friendly reference; fabricated nursing credentials (purchased diploma, license claims that do not hold up); overseas operator posing as a local remote worker; fabricated driving experience; borrowed identity (real person fronts the hiring process, someone else does the work).

Holdout set after later fixes (24 samples)

After reading the holdout misses we changed three rules (a dated bullet hid one anachronism, ship-to addresses in another state, and camera wording), then tightened several rules after a second outside review to cut false alarms (engine 1.3). These numbers are no longer a clean holdout.

Counted as flagged when the report says	Fakes caught	Honest candidates flagged
Verify first	1 of 11	0 of 13
Verify first, or two or more Worth verifying	3 of 11	0 of 13
Any finding shown	5 of 11	0 of 13

What this means: on the independent set it caught only a minority of fakes, and it raised no false alarms on honest candidates offline. With lookups on, which is the default, one honest holdout candidate (1 of 13) and one honest development candidate (1 of 22) were flagged because their made-up domains do not exist. Real typos and brand new domains will trigger the same finding. Several holdout fakes are undetectable from documents by design (a friendly reference, a real person renting out their identity). The tool is good at catching careless or document-level fakes. It cannot catch a well-made fake identity from documents alone, which is why every report ends with live verification steps. Treat a clean report as "nothing obvious", not as "verified".

Limits

Scanned image PDFs have no text to read. Paste the text instead.
Hidden white-colored text at normal size is not detected yet, only tiny text.
US phone numbers are rarely marked as VoIP in public numbering data, so most US VoIP numbers will not be flagged. Call the number live.
LinkedIn cannot be checked automatically. Use the ownership check.
Public lookups can fail or be rate limited; the report lists anything not checked.

Legal framing

Candidate Check runs in your browser on what the candidate gave you. LumeGrid does not receive candidate data and does not furnish reports, and the tool does not search criminal, credit or court records. LumeGrid does not intend to act as a consumer reporting agency under the FCRA. Whether a particular use is covered is a legal question for your counsel. Use findings only as prompts to verify, apply the same checks to every candidate for a role, and talk to your counsel about state and local rules on automated tools in hiring. This is not legal advice.