Refolk
StandardRecruiting and sourcing

The Pre-Contact Profile Authenticity Standard: Genuine, Step-Up, or Synthetic

Run any sourced profile through six cross-platform checks, grade it Genuine, Step-Up, or Synthetic with a logged reason, and never use a signal that tracks national origin.

22 min readLast reviewed November 4, 2026Read as Markdown

Key takeaways

  • GitHub writes each repository's created_at and pushed_at and each event's timestamp itself, while commit dates are whatever the client supplies, so you catch backdated history by comparing the two.
  • A Synthetic grade requires two hard indicators from two different checks, because the FBI warns that screening on isolated indicators produces false positives.
  • At LinkedIn's published 1% false positive rate for GAN faces, screening 1,000 genuine profiles would flag about 10 real people, so an AI-face flag sends a case to Step-Up and never to Synthetic.
  • Refolk's index holds 21,274 US Technical Recruiters and 4,919 people with sourcer titles, about 4.3 recruiters per sourcer, so one sourcer's authenticity grade feeds several recruiters' pipelines.
  • Name style, language quality, accent, location, and country of education are excluded from the standard because each tracks national origin and none of them tells a persona from a real person.
  • No published source measures how many outbound-sourced profiles are fake, so a team's own logged grades are the only base rate it can trust.

This standard is for sourcers, in-house recruiters, talent leaders, and founders who find candidates on GitHub, LinkedIn, and the open web. Before the first message, they need to decide whether a profile belongs to one real, consistent person. It gives you six cross-platform checks, three grades with rules tight enough that two sourcers reach the same answer, a log format, and a handoff to interview-stage identity controls. It also leaves out, on purpose, every signal that tracks national origin.

What does this standard decide, and for whom?

It answers one question before first contact: is this sourced profile Genuine, does it need Step-Up identity checks later in the process, or is it Synthetic and not to be contacted? It covers outbound sourcing, where you chose the profile. It does not cover inbound applications.

The three grades mean the same thing everywhere in this guide:

  • Genuine: the profile is consistent across platforms, rests on at least one trace the candidate does not control, and shows no unresolved indicator of fabrication. Proceed normally.
  • Step-Up: the profile is probably real, but something is unresolved or thin. Outreach goes ahead, and the case is flagged so the interview and onboarding stages apply layered identity checks.
  • Synthetic: at least two hard indicators from different checks show the persona is fabricated or borrowed. Do not contact. Log it and escalate.

A persona is the identity a profile presents: name, face, employer history, and code. A synthetic persona is one that does not belong to the person operating it, whether it was invented outright or borrowed from a real person.

Why the gate belongs before outreach

Almost every published control comes later in the process: ID document checks, liveness tests, shipping a laptop only to the address on the ID. Those controls assume the candidate came to you. Outbound sourcing works the other way. Microsoft's Jasper Sleet report describes North Korean IT workers who tailor personas to job postings and build GitHub footprints to show prior work. These personas are built to be found, so a sourcer's search query is exactly what they are designed to match.

The problem is real, but nobody has measured how much of it reaches outbound sourcing:

  • A Gartner survey of 3,000 job candidates found 6% admitted to taking part in interview fraud.
  • Gartner has forecast that one in four candidate profiles worldwide will be fake. That is a forecast, not a measured rate.
  • Checkr found 40% of surveyed US hiring managers had interviewed a candidate later found to be using a fake identity.
  • Pindrop tied about 100 of 827 applications to one developer job to fake identities. Those were inbound applications.
  • GitHub's analysis describes one North Korean team that built at least 135 synthetic identities, which reached at least 48 private codebases.

No published source separates inbound applicants from outbound-sourced profiles, so that rate is not established. You can measure your own by logging every grade this standard produces.

Who runs it

4.3
US Technical Recruiters per sourcer-titled person
Derived from Refolk's index - 21,274 Technical Recruiters against 4,919 Technical Sourcers, Talent Sourcers, and Sourcing Specialists in the United States.

The sourcer is the narrowest point in the funnel. In Refolk's index of professional profiles, there are about 4.3 US Technical Recruiters for every person with a sourcer title. One sourcer's grade feeds several recruiters' pipelines, so a mistake made at this point spreads, and a check made here covers a lot of downstream work for little effort.

Which evidence can a fabricated profile not fake?

The records a platform writes itself. On GitHub, that means each repository's created_at and pushed_at, plus the created_at on each public event, which GitHub's documentation describes as the date and time when the event was triggered. Anything the client supplies can be set to any value, including commit dates and the contribution graph built from them.

Git stores whatever date it is given. Setting GIT_AUTHOR_DATE and GIT_COMMITTER_DATE before committing is enough to backdate a commit, and Git does not verify timestamps. GitHub's own REST endpoint for creating a commit also accepts an author and committer date supplied by the caller. Open-source backdating scripts are freely available, and one of them randomly skips about 20% of days so the resulting graph looks natural. On its own, a green contribution graph with realistic gaps proves nothing.

Who wrote the evidence

  1. Self-asserted claims
    Bio, title, employer, photo, resume - anyone can type them
  2. Client-set history
    Commit author and committer dates, the contribution graph - Git does not verify them
  3. Third-party traces
    Pull requests merged by other maintainers, credits in other people's projects
  4. Server-written records
    Repo created_at and pushed_at, event timestamps - the client cannot set them
Weight each piece of evidence by who wrote it - the further in you go, the less a persona builder controls it.
The contribution graph trusts the client. The repo record and the event log do not.

Three limits you should know before relying on this:

  • GitHub publishes no official procedure for comparing commit dates against server records. The method in this guide is a practitioner method.
  • Public event data only goes back 90 days or 300 events. For older history, the repo created_at is your anchor.
  • GitHub's Users endpoint returns an account creation date. Confirm the exact field in GitHub's current REST reference before you script against it.

LinkedIn and the open web have nothing like a server-written repo date that a sourcer can query. On those platforms, the strongest evidence is whether the profile agrees with the GitHub record and what traces other people have left.

The six checks: what each proves and how it lies

Each check answers one question, and each has a known way of giving a false result. Read the third column before you grade anything.

CheckWhat it provesHow it lies
Server-side datesAccount and repos existed when claimedCareer changers open new accounts
History forensicsHistory was not backfilledMigrations, rebases, squash merges
Identity coherenceOne person behind every accountName changes, stale profiles
Footprint depthOther people accepted the workPrivate-repo careers leave little trace
Photo reuseFace is not borrowed or generatedReal people use AI headshots
Persona-network overlapProfile is not one of a setCommon name-based email formats

Identity coherence is where Microsoft's analysts begin: gather all known external accounts operated by the alias or persona, then compare them. Name, employer, dates, and links should agree. The FBI lists "claiming ownership of other people's professional networking profiles" as an indicator. That is the strong form of a mismatch: the LinkedIn profile the candidate links to verifiably belongs to someone else.

Footprint depth counts traces the candidate does not control, such as pull requests merged into repositories owned by other people, or a colleague's project that credits them. Stars do not count. A research paper titled "Six Million (Suspected) Fake Stars on GitHub", which introduced the StarScout tool, is a reminder that popularity signals can be bought.

Persona-network overlap is the check GitHub's own analysis relies on most: identical resumes with different names, reused or swapped repositories, overlapping phone numbers and email structures, and unknown employers that keep appearing. Before contact, you can only compare against what you already hold: your ATS, past logs, and other profiles in the same search. Microsoft found a public repo containing real and AI-enhanced worker images alongside LinkedIn and GitHub accounts, and saw the same AI-generated photo reused across several personas. A face that appears under two names is a strong lead.

How accurate are the detection tools?

MethodCatch rateFalse flagsSample
LinkedIn GAN-face detector99.6%1%100,000 real / 41,500 synthetic
commit-history-forensics repo2/2 fakes0/25 genuine27 repos, self-published
Reverse-image searchnot publishednot publishedn/a

The evidence is thin, and the standard is built around that fact. TPR (true positive rate) is the share of fakes a detector catches, and FPR (false positive rate) is the share of real cases it wrongly flags. LinkedIn's research detector, which only covers StyleGAN-family faces, reports 99.6% TPR at 1% FPR. Other image generators are outside its scope. The only commit-forensics figures come from a single self-published repository, which makes them a demo, not a validation. Reverse-image search has no published accuracy figure at all. Every tool result is one input to a graded judgement, never the verdict.

Grading rules: Genuine, Step-Up, or Synthetic

Grade by counting logged hard and soft indicators, not by gut feel. Synthetic needs two hard indicators from two different checks. Anything unresolved that falls short of that is Step-Up. Genuine needs every check logged with nothing open, plus at least one independent trace.

Hard indicators

  • A photo, repository, or professional profile on the persona verifiably belongs to a different, identifiable person.
  • Persona-network overlap: the same resume under another name, the same phone number or email as another persona, or repositories swapped between personas.
  • Backdated history: commits dated before their repo's created_at that arrived in single-day pushes, with no migration source found after you checked the README, the original remote, and the old host.

Soft indicators (the grade can be no worse than Step-Up)

  • An AI-face flag from a detector.
  • A reverse-image hit that cannot be tied to another named person.
  • A new or sparse account.
  • No independent traces.
  • A cross-platform mismatch with a plausible explanation you have not yet confirmed.
  • Commits older than the repo where a migration is plausible but unconfirmed.

Contradictions against independent traces

Hard contradiction foundNo hard contradiction
Contradicted and thin
Synthetic only if a second hard indicator from another check confirms it
Contradicted but deep
Step-Up - look for a migration, rename, or other legitimate cause
Clean but thin
Step-Up - never Synthetic on account age or sparseness alone
Clean and deep
Genuine - proceed to outreach
No independent tracesSeveral independent traces
A hard contradiction on a profile with deep independent traces usually has an innocent explanation, and a clean but thin profile is a new developer until proven otherwise.

Two sourcers grading the same profile should reach the same grade, because the rules count logged facts rather than impressions. When they disagree, one of them logged something the other missed, or a rule is ambiguous. Both can be fixed.

In practice, Synthetic means: do not contact, keep the log, and send it to whoever owns security incidents at your company. Do not message the persona to "test" it.

How to run the pre-contact check

Run the eight steps in order. The only reason to stop early is to record a Synthetic grade once two hard indicators from different checks are logged. No source publishes how long each step takes, so time a pilot before you set throughput targets.

Pre-contact authenticity check

  1. Snapshot the profile
    Capture every profile URL you found for the person and the date, and archive the pages. Done means an archived record with a log ID.
  2. Pull server-side dates
    Record the GitHub account creation date, each repo's created_at and pushed_at, and the public events next to the claimed career start. Note which window you saw, since events cover only 90 days or 300 events.
  3. Run history forensics
    Flag commits dated earlier than their repo's created_at and check whether they arrived in single-day pushes. Done means a count of flagged commits per repo, with any migration source checked.
  4. Check identity coherence
    Gather every external account the persona operates and compare name, employer, dates, and links across GitHub, LinkedIn, and the web. Log each mismatch with an explanation or mark it unresolved.
  5. Measure footprint depth
    Count traces the candidate does not control, such as pull requests merged into other people's repositories. Stars and repos the candidate owns do not count.
  6. Check photo reuse
    Run a reverse-image search and an AI-face check on the profile photo. Log the result, including a null result.
  7. Check persona-network overlap
    Compare against your ATS, past logs, and the rest of the search for shared resumes, swapped repos, matching phone or email patterns, and unknown employers that keep recurring. Note each overlap, or log none found.
  8. Grade, log, and hand off
    Apply the grading rules and write a one-sentence reason that cites the logged indicators. Route Step-Up cases to interview-stage controls, and do not contact Synthetic cases.

Step 4 takes the most time, because you are piecing one person together from accounts scattered across platforms. Refolk returns people already matched across GitHub, LinkedIn, and the open web from a plain-English request, so the account-gathering half of the step starts from a linked set instead of a blank search. The comparison and the judgement still happen by hand.

Step 5 is the other expensive one, and you can move it upstream. If you search for people who already have merged work in the projects you care about, footprint depth stops being a check you run afterwards and becomes part of how the candidate was found.

Record every case in the same format. A blank field is ambiguous. "None found" is evidence.

Authenticity log record
Log ID:
Date checked:
Grader:
Profile URLs (archived):
Claimed career start:
GitHub account created:
Repo created_at / pushed_at (per repo):
Events window seen (90 days or 300 events):
Commits dated before repo created_at (count per repo):
Migration source checked (README / original remote / old host):
Cross-platform mismatches (each, with explanation or "unresolved"):
Independent traces (count, with links):
Reverse-image result:
AI-face check result:
Persona-network overlap (what, and against which record):
Hard indicators (list, with the check each came from):
Soft indicators (list):
Grade: Genuine / Step-Up / Synthetic
Reason (one sentence citing the indicators above):
Handoff (Step-Up only): sent to ______ on ______

One record per profile. Write "none found" or "null" instead of leaving any field blank.

How this standard goes wrong: failure modes and false positives

Most failures are false positives, where a real person is graded Synthetic or pushed into Step-Up because a check gave a false result. The FBI itself cautions that aggressive screening based on isolated indicators can produce false positives. That is why no single check can produce a Synthetic grade.

  1. Commits older than the repo. A legitimate migration from GitLab or Bitbucket produces exactly this pattern. The maintainer of one open-source detection tool goes further: "Rebases, CI bots, squashed merges and legitimate empty commits produce the same shapes as a backdating script." Check the README, the original remote, or the project's history on its old host before you count it as hard.
  2. New or sparse account. One fake-account detector admits that a new developer with a sparse profile can legitimately score 0.75 or higher. Career changers, graduates, and people whose work sits in private repositories all look like this. Never grade Synthetic on account age.
  3. AI-face flag. A 1% false positive rate sounds negligible until you apply it at volume: screening 1,000 genuine profiles would flag about 10 real people (derived). The figure only covers GAN faces, and some genuine people deliberately use AI headshots. The flag can send a case to Step-Up and no further.
  4. Treating one later pass as clearance. Stand-ins now sit interviews and fingerprinting, so a Step-Up case that passes one live check has not been cleared. See the handoff section.
  5. Excluded signals creeping back in. A note in the reason field like "odd phrasing" or "location doesn't fit" turns a uniform checklist into a national-origin screen. The next section covers what is excluded.
  6. Title homonyms. A matching title is not identity evidence, and it is not even a reliable sign of role. The dataset below shows how often a sourcing title means something else.

What title homonyms look like in practice

MarketSourcer-titled peopleOff-target in sample of 25Off-target titles found
United States4,9192 (8%)Senior Buyer, Head of Cyber Exercises
United Kingdom4204 (16%)Procurement, billing, category lead, CEO
US to UK multiple11.7x (derived)n/an/a

Source: Refolk's index. Percentages and the multiple are derived.

In the United Kingdom sample, up to 16% of "Sourcing Specialist" style matches turned out to be procurement or unrelated roles. The same mistake shows up in candidate grading when a sourcer treats a matching title and employer name as proof. A title is self-asserted, the outer layer of the evidence stack. What backs it up is context: the employer's actual business, colleagues who list the same team, and work that other people merged.

Which signals are excluded to avoid national-origin bias?

The standard never uses name style, language quality, typos, accent, location, nationality, or country of education as indicators. Each of these tracks national origin, and none of them tells a persona apart from a real person. Location is logged once, at onboarding, and only to check that the shipping address matches the ID.

Excluded signals and why:

  • Unusual naming conventions. The FBI lists these among its indicators, but they track national origin directly.
  • Typos and machine-translated phrasing. The State Department notes that LLMs now produce convincing second-language profiles. Fluency fails in both directions: it neither clears a persona nor convicts a real person.
  • Location. The FBI names freelancers in Pakistan, India, and Nigeria in its reporting. Turning that into a location flag invites a disparate-treatment claim.
  • Accent. Under EEOC guidance, an employer may not base a decision on an employee's foreign accent unless the accent materially interferes with job performance. This matters at the interview handoff.
  • Foreign education. The 29 CFR 1606 regulations name foreign training or education requirements as a concern.

Two legal terms matter here. Disparate treatment means treating someone differently because of national origin. Disparate impact means a neutral-looking rule that disproportionately screens out a group without a job-related justification. The EEOC's national-origin enforcement guidance analyzes decisions under both, and it points to the Uniform Guidelines on Employee Selection Procedures (UGESP) for measuring adverse impact. That guidance has not been rescinded or replaced.

The enforcement climate is shifting. The EEOC has approved a new National Enforcement Plan that signals a move away from pursuing disparate-impact theories, and a DOJ opinion letter has called the existing guidelines unconstitutional. Disparate treatment remains fully in force. Check the current position with counsel before you adopt this standard. A written checklist, applied the same way to every profile, with a logged reason for each grade, protects you against both kinds of claim whichever way enforcement moves. I could not find any public documentation template for proving that a screen is applied uniformly and is job-related. The log record above is my proposal, not a regulator's.

Handing Step-Up cases to interview and onboarding controls

A Step-Up grade tells the next stage what to do. It is not a rejection. Outreach goes ahead, and the recruiter who owns the case applies layered identity controls at interview, offer, and onboarding. No single later check clears the case, because stand-ins now sit interviews and fingerprinting.

Where a Step-Up flag travels

  1. Pre-contact
    Sourcer grades and logs; the Step-Up flag travels with the case
  2. Interview
    Camera on throughout; compare photo ID to the person on screen
  3. Offer
    Identity in depth starts here; watch payment and banking requests
  4. Onboarding
    Ship only to the ID address; in-person drug test or fingerprinting
  5. System access
    Granted only after the background check completes
The pre-contact grade adds a layer that later controls cannot replace, because a paid stand-in can pass a live check but cannot change a server-written date.
StageWhat becomes visibleControl
InterviewID mismatch, manipulated video, camera refusalCamera pointed out a window
OfferCrypto pay requests, frequent bank changesStart identity in depth
OnboardingRemote desktop tools, avoiding in-person stepsShip to ID address; in-person checks

The interview signals also include a contact who changes depending on the time of day. The controls in this table come from FBI guidance. That guidance also says to require more documentation if a candidate asks for equipment to go to an address other than the one on their ID, and not to grant system access until the background check is complete. Later FBI guidance calls this approach "identity in depth", and one analysis of it argues the place to start is the offer stage. This standard adds an earlier layer before contact. It does not replace any of these.

Why one passed check is not enough: FBI PSA I-101326-PSA reports that stand-ins sit interviews and fingerprinting, and workers have bought accounts from identity-verification providers. In-person checks can now be rented. Pre-contact public evidence is a layer a paid stand-in does not cover. The FBI also says the regime keeps about 85 to 90% of what these workers earn, so the motive behind building good personas is substantial.

Fear of losing good candidates is a weak argument against in-person steps. In a Gartner survey, 62% of candidates said they were more likely to apply if in-person interviews were required.

Step-Up handoff note
Log ID:
Grade: Step-Up
Unresolved indicators (with the check each came from):
-
What would resolve them:
-
Controls requested at interview: camera on throughout; photo ID compared to the person on camera
Controls requested at offer: identity in depth; any payment or bank-change requests escalated
Controls requested at onboarding: equipment ships only to the ID address; in-person drug test or fingerprinting where available; no system access until the background check is complete
Not to be considered: name style, language, accent, location, nationality, country of education
Sourcer:
Date:

Paste into the ATS record. State logged facts and indicators only, never impressions, location, or language.

Keeping the standard calibrated and current

A standard is only as good as the agreement between its graders. Have two sourcers grade a shared batch of profiles independently on a regular cycle, rewrite any rule they disagreed on, and re-check the parts that change: the GitHub events window, detector accuracy, official advisories, and the enforcement climate.

How to keep it honest:

  • Double-grade. Two sourcers grade the same profiles blind, then compare logs. Every disagreement points to a missing log field or an ambiguous rule.
  • Time the pilot. Nobody publishes how long each step takes. Measure it on your own searches before promising throughput to hiring managers.
  • Count your own base rate. Track what share of logged grades come out Step-Up and Synthetic. No public source measures this for outbound sourcing, so your log is the only number you will have.
  • Re-check the platform mechanics. Confirm the events cap and the account-creation field against GitHub's current REST reference before changing any script.
  • Re-check the detectors. Watch LinkedIn's trust and safety engineering publications, including its work with Professor Hany Farid, for accuracy figures that cover generators beyond GANs.
  • Re-read the advisories. FBI public service announcements, State Department alerts, Microsoft threat reports, and GitHub's own analyses are where new persona tactics appear first.
  • Re-check the legal climate with counsel whenever the EEOC or DOJ changes its enforcement position.

Comparing rules with peers who already run fraud checks shortens calibration. Refolk can find technical sourcers at US software companies who mention candidate fraud or identity verification in their profile, which gives you a short list of people to compare notes with.

Before you send the first message

  • Every profile URL is archived with a log ID, a date, and a grader name.
  • Server-side account and repo dates are recorded next to the claimed career start.
  • The events window you saw (90 days or 300 events) is noted in the log.
  • Every commit dated before its repo's created_at is counted, and a migration source was checked.
  • Each cross-platform mismatch is logged as explained or unresolved.
  • Independent traces are counted, excluding stars and repos the candidate owns.
  • Reverse-image and AI-face results are logged, including null results.
  • Persona-network overlap was checked against the ATS, past logs, and the current search.
  • No excluded signal (name style, language, accent, location, nationality, country of education) appears anywhere in the reason.
  • A Synthetic grade cites two hard indicators from two different checks.
  • Every Step-Up case has a handoff note naming the controls requested at interview, offer, and onboarding.

Questions practitioners ask

How do I spot a fake candidate profile before outreach?

Compare what the platform recorded with what the profile claims. Check the GitHub account and repo creation dates against the claimed career, flag commits dated before their repo existed, and confirm the identity matches across GitHub, LinkedIn, and the web. Count work that other people accepted, check the photo for reuse, and look for overlap with other personas. Grade the profile on the logged results, not on how it feels.

Can a GitHub contribution history be faked?

Yes. Commit dates are set by the client, and Git does not check them, so a contribution graph can be backfilled to cover any span of years. Some scripts skip about 20% of days so the graph looks natural. The records GitHub writes itself are harder to fake: each repository's created_at and pushed_at, and the timestamps on push events. Commits that claim a date before their repo existed are the thing to look for.

How do I screen for fake candidates without discriminating on national origin?

Run the same logged checks on every sourced profile, and leave out any signal that tracks national origin: name style, language quality, typos, accent, location, nationality, and country of education. The EEOC's national-origin guidance covers both disparate treatment and disparate impact. Disparate treatment is fully in force even as enforcement priorities change. A uniform, written checklist with a reason recorded for each grade gives you a defence against both kinds of claim.

Should I reject a candidate whose photo is flagged as AI-generated?

No. An AI-face flag is a soft indicator, so the most it can do is make the case Step-Up. LinkedIn's research detector wrongly flags 1% of real photos, which is about 10 people in every 1,000 genuine profiles. Its accuracy figures cover GAN faces only, and some real people use AI headshots. A photo only becomes a hard indicator when you can show it belongs to a different, identifiable person.

Does passing a video interview clear a Step-Up flag?

No. FBI guidance reports that paid stand-ins now sit interviews and even fingerprinting, so passing one live check does not prove the persona belongs to the person you will employ. Keep the Step-Up flag on the case through offer and onboarding. Ship equipment only to the address on the ID, use in-person steps where you can, and grant no system access until the background check is complete.

How common are fake profiles among sourced candidates?

No one has published that number. One Gartner survey found 6% of candidates admitted to interview fraud. Checkr found 40% of US hiring managers had interviewed someone later found to be using a fake identity. Neither source separates inbound applicants from outbound-sourced profiles, and the one-in-four figure that gets quoted is a Gartner forecast, not a measurement. Your own log of grades is the only outbound base rate you can rely on.

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next