Refolk
TeardownMarket and talent intelligence

The Rival Lab Roster Teardown: Bylines to a Dated, Clustered Researcher Map

Turn one competitor's bylines into a deduplicated, dated roster with current, departed, or unresolved status, research clusters, and a key-author concentration figure.

22 min readLast reviewed November 6, 2026Read as Markdown

Key takeaways

  • OpenAlex's affiliation matcher names exactly the right institutions for 89% of random affiliation strings, and ROR excludes internal company units, so raw affiliation strings, not institution IDs, are the primary pull for a corporate research lab.
  • Match the company at the authorship level, never the paper level: a paper-level match pulls in every academic co-author and inflates the roster with people who never worked there.
  • About 92% of OpenAlex papers carry at least one ORCID but only 7.5% of author profiles do, so key authors are well identified and the long tail of interns and one-off co-authors is where split and merged identities concentrate.
  • Silence is not departure: in one Gulf-states study, at most 16 of 75 ORCID-linked disappearance cases per country showed foreign employment, so a silent researcher stays unresolved until a byline at a new employer appears.
  • Key-author concentration, adapted from the Truck Factor's 50% orphan threshold, is the smallest number of people whose removal leaves more than half of a cluster's papers with no company author, and it is an adaptation, not a published bibliometric standard.
  • In Refolk's index the UK has 257 machine-learning research scientists against 2,546 in the US, and 17 of a 25-profile UK sample sat at Meta, Google, or Google DeepMind, so one rival lab can be a visible share of a national pool.

This guide is for strategy teams and talent-intelligence analysts who have to answer three questions about a competitor: who actually does its research, what each group works on, and who has already left. The only evidence is the company's papers and preprints. I carry one rival lab through the whole job, from raw affiliation strings to a deduplicated roster where each person has a status, a last-seen date, a research cluster, and a place in a key-author concentration figure. Every number on the finished map traces back to a byline, and I mark each point where that chain tends to break.

What the finished roster contains

The finished roster has one row for each real person who has put the competitor's name on published work. Each row carries a relationship tag, a last-seen byline date, a status of current, departed, or unresolved, and a research cluster. On top of the rows sit two summaries per cluster: a member count and a key-author concentration figure.

The whole method rests on three terms:

  • Byline: the list of authors on a paper or preprint.
  • Authorship: one author on one paper, plus the affiliation that author gave on that paper. You match on authorships, not on papers.
  • Raw affiliation string: the affiliation text exactly as the author printed it, before any database maps it to an institution.

What the roster proves: a named person gave the company as their affiliation on dated, public work. What it cannot prove: who works on unpublished or internal-only projects, who stopped publishing but stayed, or who listed the company during a short visit. Each row needs evidence fields so a reader can tell which of these cases applies.

Treat headcounts as softer than concentration figures. About 92% of OpenAlex articles carry at least one ORCID (a persistent researcher identifier), but only 7.5% of author profiles do. That gap means ORCIDs cluster among prolific and lead authors. The people who drive a cluster's output are well identified. The long tail of interns and one-off co-authors is where duplicate and merged identities pile up. So a figure built from the top authors deserves more trust than a total that counts the tail.

7.5%
of OpenAlex author profiles carry an ORCID, though 92% of papers have at least one
Identity quality is best for the prolific authors who matter most and worst in the long tail, which is why headcounts wobble more than concentration figures.

Set up the roster file and the stage log before you pull anything. The stage log is where your own intermediate counts go. A roster nobody can audit later is an opinion.

Roster row schema and dated stage log
ROSTER
person_id,display_name,openalex_ids,orcid,raw_strings_seen,company_papers,total_papers,email_domains,relationship,last_company_byline,first_new_employer_byline,new_employer,orcid_end_date,status,cluster,merge_note

STAGE LOG (run date: YYYY-MM-DD)
stage,count,note
works_raw_string_pull,,
works_institution_id_pull,,
works_only_in_institution_pull,,
works_after_dedup,,
authorships_total,,
authorships_with_company_string,,
people_after_dedup,,
merges_logged,,
employees,,
collaborators,,
unclear,,
current,,
departed,,
unresolved,,
clusters,,

Paste into a spreadsheet; fill one roster row per person and one stage-log count per step, with the run date at the top.

The worked case: one rival's London lab against the UK pool

The worked case is Google DeepMind's London research group, treated as a competitor whose roster has to be rebuilt from bylines. Before you pull a single paper, size the national pool the lab hires from. That tells you whether the roster will be a footnote or a material share of the field.

In Refolk's index of professional profiles, the pool is small and concentrated:

SegmentProfilesTop employer in 25-sampleBig-tech share of sample
US, ML skill2,546Google DeepMind (3)20% (5 of 25)
UK, ML skill257Meta (8)68% (17 of 25)
US, NLP skill167Meta (8)76% (19 of 25)

The US holds 9.9 times as many machine-learning research scientists as the UK. In the UK sample, Meta (8), Google (5), and Google DeepMind (4) together account for 17 of 25 profiles. At that density, one lab's roster can be a visible share of the national pool. That works in your favour, because people who leave a lab this size tend to show up again in new-employer bylines and profiles.

How much of the UK pool one rival holds

  1. UK ML research scientists
    257

    Refolk's index, full set

  2. Profiles sampled
    25

    shares below are sample-only

  3. At Meta, Google, or Google DeepMind
    17

    68% of the sample

  4. At Google DeepMind
    4

    the worked rival

In a 25-profile sample of Refolk's UK research-scientist set, three big-tech employers held 17 profiles and Google DeepMind alone held 4.

Two caveats apply. First, the shares come from a 25-profile sample, not the full set, so read them as a direction and not as a measurement. Second, Google and Google DeepMind appear as separate employers (5 and 4 of 25). People draw that line on their profiles, but they draw it inconsistently on bylines. That is the first fork in the teardown, and the next section takes it up.

I will not print per-stage counts for this rival. Any count I printed would go stale the next time the open graph is rebuilt: the matcher's exact-match rate moved from 73% to 89% in a single version change. Record your own counts in the stage log at every step, with a date, so the next refresh can be compared line by line.

Why does a corporate lab break the standard affiliation pull?

A corporate lab breaks the standard pull for two reasons. The registry behind institution IDs leaves out internal company units, so the lab usually resolves to its parent company. And the affiliation matcher is weakest for companies. For a company lab, raw affiliation strings are the primary pull and the institution ID is only a cross-check.

Two systems are involved. OpenAlex is an open bibliographic graph that links works, authors, and institutions. ROR (the Research Organization Registry) is the open registry of organizations that OpenAlex institutions map to. ROR's scope excludes internal units of companies by design. A lab such as "X Research Zurich" therefore resolves to its parent company, together with every product team that also publishes.

BenchmarkValueWhat it tells you
Matcher exact match, random strings89%Baseline for any string
Matcher precision, most academic institutions>98%Does not apply to company labs
ORCID metadata correctness~95%ORCID is strong, not final
Author-paper pairs with ORCID (one corpus)55.7%Nearly half lack one
Author profiles with ORCID7.5%The tail is unidentified

Read the table for its gaps. The >98% precision figure covers academia. The 89% figure covers random strings. OpenAlex itself says matching is weaker for companies and small institutes with short names or no ROR record. Nobody has published matcher accuracy for corporate research labs specifically, so assume it falls below 89% until you have read a sample of matched strings by hand.

Where a corporate lab lives in the affiliation data

  1. Parent organization record
    Every team the company has, publishing or not, rolls up here
  2. Named research lab
    Usually no registry record of its own, because company internal units are out of scope
  3. Raw affiliation string
    What the author typed, and the only place the lab name reliably appears
  4. Email domain on the paper
    A supporting signal for employee versus visitor, not a match key
The lab's identity survives only in the raw string and the email domain, because the registry layer collapses it into the parent.

The first fork: which strings count as the lab

For the worked case, sort every string variant into one of three buckets:

  • Lab strings: any string naming DeepMind. These go into the pull and can qualify a person for the roster.
  • Parent-only strings: strings naming Google without the lab. Hold these in a review bucket and keep them off the roster. Some belong to lab researchers who wrote the parent name, but many come from other parts of the company.
  • City qualifiers: "London" in the string. Treat it as a soft signal. Its presence supports a London assignment, but its absence proves nothing, because many authors leave the city out.

The wrong turn here is to treat the parent institution-ID pull as the roster. It looks complete and it is fast. It also mixes in every team the parent employs.

A second, quieter wrong turn: matching on a short company name that another company shares. Before a variant goes on the list, read its sample raw strings, so a similarly named firm does not get absorbed.

Acquired and renamed labs

ROR records more than 17,000 parent-child relationships and supports Predecessor and Successor relationship types. If your rival absorbed or renamed a lab, its papers from before the change sit under the old name, and a search on the current name will miss them. Check the rival's ROR record for Predecessor and Successor links, add each old name to the variant list, and run a raw-string search on each one.

The graph does get corrected. In one screening pass, OpenAlex removed 6,191 wrong affiliation links, roughly a quarter of one percent of about 2.5 million strings. But its Affiliation Editor is open only to member institutions, so a company that is not a member has no direct way to fix its own records. Expect that noise to stay.

Step-by-step: from bylines to a dated roster

The procedure has eight steps: resolve identifiers, pull works, explode to authorships, deduplicate people, classify relationships, date status, cluster, and compute concentration. Each step has a done-state. Do not move to the next step until you can write that step's count in the stage log.

Rival lab roster from bylines

  1. Resolve the company to identifiers
    Find the parent ROR or OpenAlex institution, any child or acquired-lab records, and every raw-string variant of the lab name. Done when each variant has a real sample string beside it.
  2. Pull works from both sources
    Query OpenAlex by raw affiliation string first and institution ID second, then add arXiv records within its rate limit. Done when you have one deduplicated work list with publication dates.
  3. Explode to authorships and keep matching authors
    Split each work into authorships and keep only authors whose own affiliation string matches the variant list. Done when every author-paper pair carries a company string.
  4. Deduplicate people
    Merge on OpenAlex author ID, then ORCID, then name plus co-author overlap. Done when there is one row per person and every merge is logged.
  5. Classify the relationship
    Use email domain, dual affiliation, repeat bylines, and later university strings to tag each person employee, collaborator, or unclear. Done when no row is untagged.
  6. Date each person's status
    Record last company byline, first new-employer byline, and any ORCID end date, then apply a stated window. Done when each person has a last-seen date and a status.
  7. Cluster the roster
    Group employees by OpenAlex topics or co-authorship communities and name each group. Done when named clusters have member counts.
  8. Compute concentration per cluster
    Remove authors greedily, most papers first, until more than 50% of the cluster's papers have no company author left. Done when each cluster has one figure traceable to byline counts.

Step 1 is a desk job of about 1 to 2 hours (an unverified estimate). These are the pulls for steps 2 and 6:

OpenAlex and arXiv pulls for one rival lab
# 1. Primary: raw affiliation string search, one line per variant
https://api.openalex.org/works?filter=raw_affiliation_strings.search:"DeepMind"&per-page=200&cursor=*&api_key=YOUR_KEY

# 2. Old or acquired lab names taken from ROR Predecessor links
https://api.openalex.org/works?filter=raw_affiliation_strings.search:"OLD_LAB_NAME"&per-page=200&cursor=*&api_key=YOUR_KEY

# 3. Cross-check only: parent institution ID (never the roster by itself)
https://api.openalex.org/works?filter=authorships.institutions.id:PARENT_INSTITUTION_ID&per-page=200&cursor=*&api_key=YOUR_KEY

# 4. Slice by date if a pull hits the paging limit
https://api.openalex.org/works?filter=raw_affiliation_strings.search:"DeepMind",from_publication_date:YYYY-MM-DD,to_publication_date:YYYY-MM-DD&per-page=200&cursor=*&api_key=YOUR_KEY

# 5. Per-person history, newest first, for status dating
https://api.openalex.org/works?filter=authorships.author.id:AUTHOR_ID&sort=publication_date:desc&api_key=YOUR_KEY

# 6. arXiv per person, max 2,000 per slice
http://export.arxiv.org/api/query?search_query=au:"Firstname Lastname"&start=0&max_results=2000&sortBy=submittedDate&sortOrder=descending

Replace the quoted strings with your own variant list and add your API key; keep arXiv calls to one every three seconds on one connection.

Step 2: pulling within the source limits

Each source has a different budget and a different idea of what a dated affiliation is:

SourceRate or budgetResult capDated affiliation
arXiv API1 request per 3 s, one connection30,000 per queryOptional field, version dates
OpenAlex APISmall free daily budget with key, 100 req/s10,000 basic pagingRaw string per authorship
ORCID employmentNot established hereNot established hereStart and end dates, self-reported

Here is the arithmetic for a full arXiv pull. A 30,000-result cap at 2,000 per slice is 15 calls, which at one call every three seconds takes at least 45 seconds. OpenAlex requires an API key, and the documented free budget is $1 of usage per day. Check the authentication docs before a large run, because budgets change. If a pull hits the 10,000-result basic paging limit, slice it by publication date as in pull 4.

Treat arXiv as your source for dates, not affiliations. Its affiliation field is optional and often empty. The arxiv.py client by Lukas Schwab exposes per-author affiliation where one exists. For conference proceedings, one study gave up on the databases and parsed PDFs directly, because of missing articles and omitted metadata. Do the same for any venue where your rival's flagship work appears and the database record looks thin.

Step 3: the authorship cut

This step is where most rosters go wrong. A paper-level match tells you the company is somewhere on the byline. It says nothing about the other authors. On a joint paper between the lab and a university, every academic co-author would come in as staff.

Published research does define "company papers" at paper level, but the definitions disagree. One counts any paper with at least one big-tech author. Another uses only the first author's affiliation. Use either definition to describe what the group publishes, never to decide who works there.

Step 4: the identity cascade

Merge people in a fixed order and write down every decision:

  1. OpenAlex author ID. This is your starting row.
  2. ORCID. Merge rows that share one. ORCID metadata is right about 95% of the time, so a shared ORCID is strong evidence but not final.
  3. Name plus co-author overlap. Merge rows with the same name and nearly the same set of co-authors.

Know which direction the graph is likely to be wrong in. OpenAlex's AER v4 rebuild split about 3.2 million overmerged author profiles, and OpenAlex reports that splits are largely done and merges are next. So for now, expect one person spread across several rows (a split identity) more often than two people fused into one (a merged identity). One audit put split-identity noise at an upper bound of 0.15% of co-author edges. A national-registry match in Brazil found 8,861 merged cases against 670,345 authors, which works out to 1.3%. Nobody has published the expected dedup loss for a single company roster.

Step 5: employee, collaborator, or unclear

Each signal proves something narrow and can mislead in a predictable way:

  • Email domain. Proves the author had a company mailbox when the paper was written. Published studies use it only to support other evidence, because on its own it does not tell you the type of employment.
  • Dual affiliation. A company string next to a university string. At least one study counts every such paper as collaborative. It misleads when an employee also holds an honorary university post.
  • Repeat bylines. The share of a person's papers that carry the company string. Nobody has published a threshold, so pick one, write it in the roster notes, and apply it consistently.
  • Later university string. A university affiliation on papers after a company-affiliated one is the classic pattern of a PhD student who visited.

When the signals conflict, tag the person "unclear". Don't break the tie in the rival's favour.

How do you date each person as current, departed, or unresolved?

Record three dates for each person: the last company-affiliated byline, the first byline at a new employer, and any ORCID employment end date. Tag someone departed only when a new-employer byline or an end date comes after their last company byline. Everyone else is current or unresolved, depending on a window you state on the page.

There is no public standard for that window. Nobody has measured month-level lag between a job move and a changed preprint affiliation. One mobility study builds in a 2-year window of publication delay, and another source notes that there is always a lag between a change in affiliation and the first paper that reveals it. I use 24 months as my convention for the line between current and unresolved, because it mirrors that published delay window. Label it as a convention wherever the roster is shown.

Status call from two kinds of evidence

Company byline inside the windowNo company byline inside the window
Current
Keep, and re-check at the next refresh
Conflict
Compare dates; overlapping strings may be a dual post, a later new-employer date is a departure
Unresolved
Hold, and check the profile side before saying anything
Departed
Record last-seen and first-new-employer dates
No new-employer evidenceNew-employer byline or ORCID end date
Departure needs positive evidence of a new employer, so a recent silence alone never moves anyone past unresolved.

Read ORCID employment end dates as evidence, not proof. The researcher or the employer enters them. Use arXiv version dates to pin down when an affiliation was actually declared, because a revised preprint can carry a different affiliation from the first version.

Bylines can only settle an unresolved person when that person publishes again, which may take more than a year. The profile side moves faster. Refolk searches LinkedIn, GitHub, and the open web in one plain-English query, so you can check the whole unresolved bucket in one pass instead of looking people up one at a time:

A profile match still needs a byline or a dated move before the status changes to departed. Write the source of the evidence into the roster's new_employer field.

A silent byline is a publication-lag artifact until a new employer shows up in print.

Clustering the roster and computing key-author concentration

Cluster the employees by OpenAlex topics or by co-authorship communities, then compute one concentration figure per cluster. The figure is the smallest number of people whose removal leaves more than half of the cluster's papers with no company author.

There are two ways to cluster:

  • OpenAlex topics. Fast, and the labels come ready-made. Topics tag papers, not people, so put each person in the topic that covers most of their company papers.
  • Co-authorship communities. Groups of people who repeatedly publish together. These reflect working groups more closely than topics do, but you have to name each cluster by hand from its papers.

The concentration figure adapts the Truck Factor, a software measure of how many people a project can lose before it stalls. The published algorithm (Avelino et al.) keeps removing the author of the most files until more than 50% of the files are orphaned. A comparison study found 50% to be the best threshold. Swapped onto papers, it runs like this:

  1. Take the cluster's papers that have at least one employee authorship.
  2. Count each employee's papers in that set.
  3. Remove the employee with the most papers. Any paper left with no company author is now orphaned.
  4. Recount and repeat until more than 50% of the cluster's papers are orphaned.
  5. The number of people removed is the cluster's concentration figure.

Expect small numbers. Across 133 popular GitHub applications, 46% had a Truck Factor of 1. Nobody has published paper-cluster equivalents, so your figure is original work. Put the method next to it.

Identity errors push this figure in opposite directions. A split identity spreads one person's papers across two rows, so neither row looks dominant and concentration reads too high, meaning the cluster looks less dependent on any one person than it is. A name collision piles a stranger's papers onto one row and makes that person look more central than they are. This is why Step 4's merge log matters more here than anywhere else.

How this goes wrong: failure modes and false positives

Most errors in a byline roster come from identity and affiliation data that looks clean but isn't. A few come from reading silence or a short visit as a fact about employment. Each failure has a check you can run before the roster leaves your desk.

Identity and affiliation failures

FailureEffect on the rosterCheck
Name collisionA stranger's papers inflate a personSudden topic or country jump; conflicting ORCIDs
Split identityOne person becomes two rowsSame ORCID or near-identical co-authors across IDs
Parent collapseProduct-team authors enterFilter on raw strings, not institution ID
Short-name over-mergeA similarly named firm gets absorbedRead raw_affiliation_string per authorship
Paper-level labelingAcademic co-authors counted as staffEnforce the Step 3 authorship cut
Acquired-lab stringsPre-change papers are missedROR Predecessor and Successor links; old-name search

Two warning signs deserve a closer look. First, an ORCID-to-author assignment is sometimes wrong, so a shared ORCID does not settle a collision on its own. Second, in some cases an ORCID has been assigned to more than 10,000 records. An ORCID attached to an implausible number of works is a collision, not a prolific author.

Status failures

  • Visiting or intern stint. A PhD student appears on 1 or 2 papers with a dual affiliation and gets tagged as an employee. The check: look at the email domain, any university string on later papers, and the share of the person's papers that carry the company string. When it misleads, it looks like a junior researcher whose company-affiliated papers all fall within a short span.
  • False departure. Someone goes quiet because of publication lag or internal-only work and gets tagged departed. The check: does a byline or a dated move at a new employer exist? If it doesn't, the status is unresolved. The Gulf-states figures above show how often silence means nothing.

Keeping the map current

A byline roster goes stale in two ways: people move, and the open graph gets rebuilt underneath you. Re-run the pull whenever you need a fresh answer, compare the new stage log against the last one, and re-run Step 4 every time OpenAlex ships an identity change.

Three things change between runs:

  • The matcher. Exact-match accuracy moved from 73% to 89% in one version, so stage counts can shift with no change at the rival. Check the OpenAlex institutions help page for the current matcher version before you compare runs.
  • Author identity. OpenAlex says merges come next. When they arrive, rows you logged as separate may consolidate, and concentration figures will move.
  • The unresolved bucket. Each refresh should move some people out of it. A bucket that only grows means your window is too short or your profile-side checks have stopped. Re-running the departure query in Refolk between paper cycles keeps the departed side current without waiting for the next preprint.

Before the map goes to leadership, verify:

Before you call the roster done

  • Every roster row lists at least one raw affiliation string that matches the variant list.
  • No person is on the roster because of a paper-level match alone.
  • Parent-only strings sit in the review bucket, not on the roster.
  • Old lab names from ROR Predecessor or Successor links are on the variant list and were searched.
  • Every merge and split decision has a line in the merge log.
  • Every departed tag cites a new-employer byline or ORCID end date that comes after the last company byline.
  • Every silent person without new-employer evidence is tagged unresolved, not departed.
  • The status window is printed on the page and labelled as an analyst convention.
  • Each cluster's concentration figure can be recomputed from the byline counts in the roster.
  • The stage log is dated and has a count for every step.

Questions practitioners ask

Can I find researchers at a company from papers alone?

Yes, for the people who publish, as long as you match on each author's own affiliation string and not on the paper as a whole. Papers will show you who put the company's name on dated public work. They will not show you internal-only staff, people who stopped publishing but stayed, or reporting lines. So label the roster "publishing researchers" and keep an unresolved status for anyone the bylines cannot place.

Why not just filter OpenAlex by the competitor's institution ID?

ROR, the registry behind OpenAlex institutions, deliberately leaves out internal units of companies, so a corporate lab usually resolves to its parent. On top of that, the matcher is weakest for companies with short names or no ROR record. An institution-ID pull therefore brings in product teams and unrelated groups. Use raw affiliation string search as the main pull and keep the institution ID only to catch strings you missed.

How long after someone leaves does their preprint affiliation change?

No published study measures that lag month by month against dated job moves. The best proxy is mobility research that allows a 2-year publication-delay window and treats papers listing only the old affiliation in that period as pre-move work. Treat any cut-off you use as your own convention, write it on the page, and only tag someone departed once a byline at a new employer or an ORCID end date proves it.

How do I tell an employee from an intern or visiting researcher on a byline?

Combine four signals: the email domain on the paper, whether the author lists a university alongside the company, how many of their papers carry the company string, and whether later papers show a university affiliation. Published studies use email domain only to support other evidence, and at least one treats any university-plus-company affiliation as collaboration. Nobody has published a repeat-byline threshold for confirming employment, so set yours and state it.

Is the key-author concentration figure a recognized bibliometric measure?

No. It adapts the Truck Factor algorithm from software research, which removes the author of the most files until more than 50% of files are orphaned. Here papers replace files, and a paper counts as orphaned when it has no company author left. One comparison study found 50% to be the best threshold, but only on code. Publish the method next to the figure so a reader can recompute it.

How many arXiv API calls does a full pull take?

At most 15 for a single query. arXiv caps a query at 30,000 results and returns at most 2,000 per slice, and its terms allow one request every three seconds over a single connection, so the full pull takes at least 45 seconds. The limit covers every machine you run, so parallel scripts break the terms. Because affiliation is an optional arXiv field, use arXiv for version dates and OpenAlex for affiliations.

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