How this gets made

Every page on VC Beast is published under a named human editor. Michael Kaufman holds the byline and the final call on what runs, what gets held, and what gets pulled. There is no anonymous editorial voice here, and no page ships that no one is willing to sign.

That is the load-bearing commitment on this page. Everything below it is process; this is the part that assigns the consequences to a person.

This is an AI-assisted publication and we would rather say so than be found out. Automated systems do the work automation is good at: gathering filings and public records, extracting structured data from documents, assembling first drafts against a research brief, and checking finished pages against our own house rules at a scale a human editor could not sustain.

What those systems do not do is decide. They do not choose which claims are strong enough to publish, which sources are credible, which ranked order is defensible, or which page is ready. Those are editorial judgments, they are made by a human, and the byline records who made them.

We describe the capability rather than the supplier deliberately. Which systems sit in the pipeline changes; the standard the output is held to does not, and that standard is what a reader is actually entitled to know.

Editorial standards that live only in someone's head decay. Ours are written as tests, and they run before anything ships — a page that violates them fails the build rather than reaching a reader and waiting to be noticed.

The gates are boring on purpose: they check the things that are easy to get wrong at volume and expensive to find later. Broken internal links, claims lacking a source, pages competing with each other for the same search intent, structured data that does not match the page it describes, and routes leaking onto a sibling property that never claimed them.

Those gates are run against fixtures that are themselves deliberately broken, because a gate nobody has watched fail is not a gate. We test that the checks catch the errors they exist to catch.

Claims are verified against primary sources — the document, the filing, or the named person — rather than against another publication's summary of them. Where a figure comes from a third party, we say which one, and we link to it where it is public.

A claim we cannot source, we do not publish. That is a real editorial cost: it kills clean-sounding sentences regularly, and it is the reason some pages are shorter than the competing ones.

We will publish errors. A publication operating at this volume that claims otherwise is either not looking or not telling you.

So we look on purpose. Alongside the automated gates we maintain deliberate checks designed to surface claims that have quietly gone stale or wrong — the kind of error that survives because nothing in the system was watching for it. When one of those checks fires, or when a reader tells us, the page is corrected and the correction is logged publicly.

Corrections are dated, describe what changed and why, and stay up permanently. We do not silently edit a published page and we do not delete a correction once the embarrassment fades. The log is at /corrections; if it is empty, it means nothing has yet needed correcting, not that the policy is decorative.

When we put tools, platforms, and firms in order, the order comes from a stated methodology on the page itself — the criteria, how they are weighted, and what would move an entry up or down. If we cannot state why something placed where it did, it does not get a rank.

Placement is not for sale. Sponsorship and affiliate relationships exist and are disclosed on the pages where they apply, and they buy exactly one thing: disclosure. They do not buy inclusion, position, or a kinder assessment, and an entry that pays is labelled as paying.

This is the commitment on this page that is worth the most money to break, which is why it is written down where a reader can hold us to it.

When we publish original data — the number of first-time funds that filed a Form D in a given quarter, for example — the number is produced by a named script against a stated dataset, not assembled by hand in a spreadsheet that no longer exists. The script, its inputs, and the date it was run are recorded, and re-running it reproduces the published figure.

That is what lets us correct a data claim properly instead of quietly adjusting it: when the underlying source changes, we re-run, we compare, and if the number moved, that is a correction and it gets logged like any other.