About & method

What this site is, who keeps it, how it decides what goes on the page, and how a mistake gets corrected.

What this is

b0b.dev is independent open-source intelligence (OSINT) research. Its core is one long report - Project Anglerfish, twenty-five sections - on the Epstein-Maxwell record and the wider architecture of surveillance, finance and institutional power, with a companion map of 1,166 locations and 455 connection lines. Every claim carries a tier; every map line carries the tier of the relationship it shows. It is also a long-running work of art in which the method is part of the work: the tiers, the published nulls and the claims killed on the page are the point, not decoration.

Who keeps it

The author and director is Gianni Arone. He sets scope and makes every contested editorial call. No institutional affiliation is claimed by this page; the work is presented as independent research and should be judged on its sources, not on a credential.

How the research is done

• Sources are fetched by one identified crawler. It obeys robots.txt, never disguises itself, never solves a challenge and never retries a refusal under another name - because a document obtained by defeating an access control has contaminated provenance. Every fetch, and every refusal, is logged with its time, status and SHA-256 hash in the source ledger.
• A refusal is a finding. Sites that block the crawler are recorded as nulls, not worked around; a claim that rests on a source the crawler could not read is held at a lower tier.
• Quotes and figures are read at the source. A search result or a summary - human or machine - never carries a quote onto the page; it is re-read in the fetched copy first.
• AI assistance is used and disclosed. Research, drafting, verification and deployment have been carried out with Claude, an AI model made by Anthropic, working under the author’s direction - including the work behind this page (October 2026). Anthropic also appears in the report as a subject; that conflict is disclosed where it arises (Sections X and XXV). The AI’s own recorded failure modes - drifting toward agreement, loaded word choices, overstatement - are among the things the method is built to catch.

The practices it applies

The report applies the working practices of intelligence analysis rather than claiming an institutional standard:

• Tiered claims - documented, attributed, labeled, contested, unsupported - with nothing promoted by repetition.
• Named sources and provenance - the claimant named for every attributed claim; the fetch ledger for every source.
• Alternative hypotheses, scored - the central conclusion is set against its rivals on the same record, and the rival it has most trouble excluding is named (Section II).
• Stated falsifiers - what would lower each major conclusion is written down before the evidence arrives (Read this first).
• Chronology and the anti-map rule - a shared timeline is not a chain; map lines are tiered so proximity cannot pass as proof.
• Symmetry - any discount applied to evidence against a hypothesis is applied to evidence for it.
• Published corrections - errors are fixed in place and listed in the corrections ledger.

Standing rules

• Living-person floor. No living person who has not been charged is asserted or implied to be a criminal or an intelligence asset. Documented affiliation is never presented as an operational tie.
• No de-anonymising. Redactions protecting living people are never reversed; residences carry street names only.
• Stop rules. Victim interviews, protective-order material and imagery-heavy documents are not read past identification.
• No numerology. Dates and numbers are recorded as facts and never read for meaning.

How a correction is decided

• A claim that a primary source contradicts is lowered or removed, whoever raised the problem - including the author’s own claims and the AI’s.
• Corrections are made in place, without narration in the text, and entered in the ledger with the date, what changed and how the error was found.
• Contested and defamation-adjacent material is held for the author’s decision rather than published automatically.
• To send a correction, use the researcher’s guide at the end of the report: the claim, the passage, and the source that contradicts it.

The data

The map and the source ledger are published as data so that the interpretation can be thrown away and the record kept: map-data.json, map.kml, the map as text and sources-ledger.json.