What we threw away.
Any model can produce a list of names. Whether they are real is a separate question, and it is answered by a deterministic layer that runs after research and calls no model at all. Every cited URL is fetched, graded against the page it returns, and dated. What fails is published here rather than patched.
Four outcomes, and they add to the total.
There is no residual bucket. Every URL behind every record on this site lands in exactly one of these, which is what makes the count checkable rather than rhetorical.
- 3,300Live92%Fetched, body readable, the named entity found on the page
- 99Bot-blocked3%Server answered and refused an automated client. Never verifies a record alone
- 192From archive5%Live page gone. Dated snapshot recovered and labelled as a snapshot
- 10Dead0%Unreachable under every signature tried, with no usable snapshot
Each URL is attempted under several browser signatures before it is called dead, then resolved against the Internet Archive. Evidence carries a 30-day time to live: past that it is re-fetched rather than reused, so nothing on this page rests on a read older than 30 days. Last full audit August 10, 2026.
48 of them could not be resolved against the archive on this pass, because archive.org rate-limited the client. They are recorded as unchecked, not as having no snapshot. Those are different facts, and collapsing them is how a verification claim quietly becomes untrue.
How each record was actually proved.
A surname on a page proves very little. Every record is graded by the strongest evidence found for it, and the grade is what ships, not the number of links found.
- Structured data38732%
A schema.org Person or Organization record. Machine-readable identity
- Page title57348%
The page is about them, not merely mentioning them
- Name beside organisation15613%
Both inside a 60-word window, which rules out a namesake
- Bare name20%
Not sufficient alone. Kept only behind a second reachable source
- No readable match756%
Survives only on corroboration across multiple reachable sources
1,193 records, graded by strongest match
- Regulatory261%
Filings and public registers
- Primary2,02752%
The organisation's own domain
- Press43811%
Named trade and financial titles
- Other1,42136%
Everything not in the tiers above
3,912 citations, graded by source authority
The corpus knows how old it is.
A map that cannot say when it was last true is not a map. Records are dated from their sources, and the three states are reported separately because 'we cannot date this' is not the same claim as 'this is old'.
- 902Dated within two years
- 197No readable date
- 94Newest source over two years old
Separately, 5 people are absent from their own firm’s live roster while their profile page is still up. Roster diffing catches that class of decay before a link ever breaks, and those records are named below.
Everything the audit flagged, by class.
Nothing on the site currently fails the bar, so this is the softer stuff: sources that are ageing, undated, recovered from an archive, or dead. Records that stop clearing the bar get named here rather than quietly removed.
Absent from their own firm's published roster. Firms leave stale profile pages up for months after someone goes, so a missing roster entry is treated as evidence against the record rather than as a neutral non-match.
- expertMitchell GoldsteinNot listed at https://arcc.ares.com/people
- expertSteve SwinneyNot listed at https://kodiakbp.com/leadership/
- expertAhmed Yahia Al IdrissiNot listed at https://www.mgx.ae/en/leadership
- companyPDS Health (Pacific Dental Services)Not listed at https://www.pdshealth.com/who-we-are/
- companyRyan Specialty Holdings, Inc.Not listed at https://ryanspecialty.com/who-we-are/
The funnel, theme by theme.
Each theme is one thesis typed in plain language and run end to end. Nothing between them was hand-curated, and the bar did not move to accommodate any of them.
Researched is what the pipeline proposed. Verified is what survived the fetch. Click a row for the thesis that produced it.
The bar, stated precisely
A fetched page identifies the entity: named in its structured data, named in its title, or named alongside its organisation inside a 60-word window. Ships.
The page resolves but cannot be read by an automated client, either JavaScript-rendered or bot-blocked. Ships only where a second independent reachable source agrees.
Cleared verification when generated and does not today. Named on this page rather than removed.
No cited URL resolves at all, live or archived. Dropped from the dataset and listed.
The verification layer is deterministic and calls no language model. It resolves URLs, parses structured data, extracts titles, matches names positionally, and writes every result to a content-addressed evidence store keyed by URL. Re-auditing the full corpus of 3,601 links against that store completes in about a second with no network traffic, which is what makes running the audit continuously rather than once affordable. Nothing here is patched to pass.