17 — Methodology

How this was reported, and what it does not claim

Publishing the method is the thing that separates this from the pages described in section 02. Those pages have no method. This one has a specific and checkable one, including its failures.

How sources were graded

Every claim on this page carries one of four marks, and each mark is a colour, a shape and a word together — never colour alone, so the page survives greyscale printing and colour-blind reading intact.

The line between Verified and Partial

This is the load-bearing decision on the page, so it is stated explicitly rather than left to the reader to infer.

Verified means a document, record, or artifact that was examined directly. A page fetched over HTTP and read. A dated Internet Archive capture. A government filing. A video in which the subject says the words himself. Things a reader can go and look at.

An outlet reporting what an official told a reporter is graded Partial — regardless of the outlet's reputation, and regardless of how many outlets carry it. This is not a judgement about any newsroom's honesty. It is a statement about what can be checked. When the underlying document is not published, the reader is trusting a chain they cannot inspect, and outlets repeating one another does not lengthen that chain — several outlets may be relaying the same office, the same spokesperson, the same phone call. Breadth of coverage measures travel, not verification. That is the same error this piece identifies in the content farms; it simply occurs at a more respectable tier.

Applied consistently, this rule costs the piece its most convenient conclusions. The officially reported cause of death is not in the verified column. Neither is the fact of the death itself. Both are graded Partial: reported, coherent, probably true, undocumented. If a death certificate, an autopsy report, or a direct statement from the Paulding County Coroner or the GBI is released, those claims move to Verified and this page will say so with a date.

Where a number appears beside a mark it is a 1–10 confidence score: 10 a document or artifact examined directly · 7–8 strong primary, or a record seen at one remove · 5–6 named press relaying an unpublished source, or single/partial sourcing · 3–4 self-reported or aggregator · 1–2 content farm or anonymous rumour. A claim graded 2 never renders like a claim graded 9.

Said versus done — and who did the checking

Two distinctions run underneath every grade on this page, and both were applied late, after the first version got them wrong.

A promise is not a delivery. The first version of this page reported that the archive "had been rolling out, free, for nine and a half months" and marked it Verified. What was actually verified was a post saying it would be. Whether any complete text was ever obtainable free is unestablished, and the page now says so. A claim about what someone said can never inherit the grade of a claim about what was done.

A research pass is not an examination. This piece was built on two multi-agent research rounds, and their findings are good — but an agent reporting that it read a document is not the same as the document having been read here. Where that distinction mattered it has been applied, and it cost the page real claims: a corporate filing was withdrawn from the Verified column when it could not be reached here, and the absence of Augusta press coverage was downgraded when the search could not be reproduced. What sits at Verified is what was opened and read at this desk: the live and archived FelzBooks pages, the felzinvestment redirect, the Richmond County court docket, YouTube's own channel feed, the Georgia corporate record, and the two 7 August articles.

Where a Verified claim rests on the research pass rather than a direct examination, it is either downgraded or its basis is stated in the entry itself. Readers should assume that anything not described as read here was not read here.

What was actually done

Two multi-agent research passes covering media triage, the primary digital footprint, family statements, cause-of-death provenance, business and financial records, the official-records base rate, monetisation forensics, a Spanish-language Costa Rican sweep, and account stewardship — roughly 730 source fetches across 22 agents, with adversarial verification passes that argued against the findings rather than for them. Then a pre-publication verification round on this desk: live HTTP fetches of every site named, the Georgia corporate record, the App Store listings, the comparative Costa Rican case, and the two reports of the official determination.

Where a document could not be reached directly it is said so. El Observador returned HTTP 403 to direct fetch and was read via published excerpt. Instagram and Facebook are behind login walls; their post timestamps were decoded from post IDs. The Wayback Machine has never captured his Instagram or Facebook, so no claim about deletions or edits to those accounts can be checked in either direction, and none is made.

What changed during the writing, and why

This piece was researched while the cause of death was genuinely undetermined. The original plan was to lead on that vacuum and to present the circulating cause-of-death claim as a case study in content-farm provenance, since at the time it had no sourcing of any kind. The August 7 reports landed mid-build. That framing became false, and it was replaced rather than softened. The verification log recording this decision is part of the file for this story.

What this piece deliberately does not claim

• It does not claim the official determination is wrong. Two outlets, two named authorities, in agreement, is meaningful.
• It does not claim the determination is established either. No document has been released and neither office has spoken publicly, so it is graded Partial rather than Verified.
• It does not claim, as established fact, that Joe Felz is dead. That will read as strange, and it is deliberate. The death is reported by his partner, treated as real by people who knew him, uncontradicted, and probably true — but not one documentary artifact supports it: no death certificate, no obituary of record, no funeral, no statement from any office. A piece whose whole subject is the distance between "reported" and "established" cannot quietly close that distance for its own central fact.
• It does not name a place or date of death, because no source establishes one.
• It does not assert that any living person faked a death, committed fraud, or profited deliberately. Where money is moving with no name attached, that is stated as an open question, which is what it is.
• It does not assert a cause for the gambling-domain redirect.
• It does not name Brandy beyond the first name in the primary post, or assert her relationship to anyone.
• It names Sarah Felz as the other party to his 2022 divorce, sourced to the Richmond County Superior Court Clerk's own docket, read directly here. That is a neutral public record, and Felz discussed the divorce publicly himself without naming her. It does not claim the court record and the influencer are documented to be the same Joseph Felz — dockets carry no addresses or dates of birth, so that link is a strong inference and is labelled as one.
• The same docket search returned four other Felz family matters, and on the editor's decision all of them are published, in a table that gives each disposition the same weight as its filing. Two were petitions that ended with no adjudicated finding against anyone — one dismissed without prejudice — and one modification remains open. They are published because they are public records verified at the issuing authority, and because printing only the docket that suited the narrative would be the selective sourcing this piece exists to criticise. The page asserts no causal link between any of them and his death, because the record supports none.
• It does not name his children. He said publicly, in his own words on a January 2024 podcast, that he had two — that much is his own statement and appears on the page. Their names do not, for three reasons, in ascending order of importance: the names in circulation are wrong (one surname is a garbled version of their grandmother's maiden name, lifted from their grandfather's obituary); the only near-primary attribution is one ambiguous parenthetical in a translated, staff-bylined outlet; and they are minors in a story about a suicide, whose father never named them on any of his own channels across years of daily posting. Publishing them would mean repeating a fabrication or correcting one — and correcting it would put two real children's names into this story on the authority of a document that was never about them.
• It does not reproduce method detail beyond the officially reported determination, and includes no scene detail of any kind.
• The photographs on this page are not photographs. Every image here is AI-generated, and every one depicts a place, an object, or a texture — never a person. None depicts Joe Felz, any other real individual, or the death scene. That line is held for the piece's own protection: on a story about fabricated material circulating as real, a synthetic photograph of the man whose death is in question is the single asset most likely to be screenshotted, stripped of its caption, and recirculated as evidence. They are atmosphere only — nothing in them is evidence, and nothing in them should be read as a depiction of a real place in this story.
• The cover art in the library is generated too. Seven titles on the top shelf carry a made image of a period binding — leather, cloth, gold foil, the real title of the real edition. They stand in for the editions the way the atmosphere images stand in for places, and the rooms behind them are invented. No cover depicts a person, and none is a photograph of any particular copy. Every other book on those shelves is drawn in CSS rather than pictured at all.
• Each entry in the library states the evidence for its own inclusion — what he said about that title, where, and how well attested it is — and carries a link to the edition it was taken from. Titles he named that are not on the shelf are listed as missing rather than quietly replaced by something similar.
• Machine-written prose is fenced, named, and graded. The version-diff section runs a local model over the aligned texts and prints its description of what differs between editions. That prose appears only inside a marked block carrying the model's name and an Unverified badge, because a model that has read nothing cannot clear the bar the rest of this page is held to. It is a lead for a human to check, never a finding. The verse text beside it is OCR of the scans, which can be wrong, and is corrected on report like anything else here.

Safe reporting

That coverage of a suicide moves the rate is not seriously contested; which direction it moves depends on how the coverage is written. That makes the how a matter of consequence rather than taste. This page follows the Recommendations for Reporting on Suicide and the World Health Organization's guidance for media, and the restraints they impose are visible in the text rather than promised here.

No method detail beyond the officially reported determination, and no scene detail of any kind. Where a source supplied either, it was dropped rather than paraphrased — a paraphrase of method is still method. No photograph of the man, and no generated image of him either, for the reason given above. Nothing here presents the death as a response to any circumstance the page names: the archive, the money, the docket and the network appear because the public record about them was falsified, not because they are being offered as an explanation. This piece takes no position on why, because it has no evidence about why — and a story assembled out of financial pressure and a death makes that argument whether or not it intends to.

That last restraint is the expensive one. The strongest narrative available here — a man under pressure, a business coming apart, a death — is precisely the one the record does not support, and declining to imply it is most of the editorial work on this page.

Crisis and bereavement resources sit in the footer of every page of this investigation rather than at the end of this one, so that they travel with whichever page a reader actually lands on. If you have come to this looking for somebody to talk to, that block is at the bottom of this page too. If you have come to it having lost someone, the last line of it is the one for you.

18 — Sources

Sources

Grouped by tier. Content farms are listed because provenance is part of the story; they are cited as evidence of what was published and when, never as evidence of what is true.