Brief: Wikipedia as an anchor, read as a fractal semantic graph

Brief: Wikipedia as an anchor, read as a fractal semantic graph

25 October 2026. Written by ontologist@php.newsroom from the editor's direction in its session on 24 October. For

the ontologist (the section), the researcher and the collectors (the references), and the fact-check desk (the

statements). Board rows at the end.

The direction

The editor, 24 October, after asking what the site held of https://en.wikipedia.org/wiki/PHP2Shell (one stale

capture, an island in the graph, half its references uncaptured):

• Wikipedia is a trusted source, "which in fact we should be able to eventually prove, from the number of issues vs

other publications".

• Wikipedia is "one of the biggest sources of truth that we have", and may already hold the data "in a kind of

quasi-graph format" (DBpedia was named).

• Take "a very semantic fractal graph approach": a complete, separate section, "a whole folder and subfolders and all

sorts of data"; "regular calls to Wikipedia", capturing the delta.

• This is the chance "to show the next generation of Wikipedia": Wikipedia "is hard to consume, it's too much, you

can't diff easily, you can't see what you've seen already, you can't follow the links".

• A brief for the researcher: "the researcher should have been following every one of those links", because

Wikipedia's quotes of those pages are quotes this site can make, once the pages are captured.

• The view wanted: Wikipedia "based on facts and opinions. Wikipedia is usually very good because you should not have

opinions; most of it should be facts. So we should be linking those facts with the evidence. And then that becomes

one of the anchors that we have on this website."

What exists now (v0.16, 25 October)

• admin/wikipedia/fetch.py: every revision of an article (364 for PHP2Shell, from 07:50 UTC on 21 October), with

its wikitext, time, editor, size and summary; the Wikidata item (Q141801185) with its claims and their labels; the

198 articles that link in. Written to data/wikipedia/<article>/. Anonymous editors are recorded as anonymous,

never by address. One request every two seconds, with the API's own pauses respected.

• admin/build/wikitext.py: the parser. One statement per sentence, its references attached, its class read from

its wording: record, stated, attributed (a named party said, reported, estimated), quote, hedged. No model.

• /wikipedia/, /wikipedia/php2shell/ (the article as a graph, the Wikidata item, what changed since the vault's

capture, the counts a measure of trust would start from), /statements/ (every statement with its evidence),

/references/ (every reference: captured here or not, its publisher as a party, the claims of this site on the

same page), /history/ (every edit as the statements it added, removed or reworded).

• data/wikipedia/php2shell/graph.json: the article in the ontology's new wikipedia scope (article, section,

statement, reference, revision, editor, item; verbs with inverses), joined to the story graph by captured_as.

From the story graph, the article's source node and every source a reference is captured as zoom into it.

The model: where Wikipedia's facts and this site's meet

• A statement there and a claim here are never declared "the same". They meet at the captured page both cite:

statement cites reference, reference captured_as source, source cited_by claim. The shared evidence is the

join (graphs.sgit.ai: join through the intermediate layer, never paragraph to paragraph).

• A class is read from wording, never from truth. Wikipedia's rule (state facts, attribute opinions) makes the

wording informative: attributed and quote are where the article hands a view to a party; record and stated

are what it asserts itself; hedged is where it is being careful. The fact-check desk grades; this section counts.

• Trust is to be shown, not assumed. The counts kept per article: statements cited, references captured and how

many are primary, reverts per revision, contradictions of this site touching the shared evidence. A score is not

computed; the comparison against other publications needs the same counts for them (board).

DBpedia and Wikidata

Wikidata is the live structured sister of Wikipedia and already holds the item: two CVE ids, the discoverers (Wiz,

OpenAI, Calif), "time of discovery" August 2026, "announcement date" 20 October 2026, links in 28 languages. Those

dated claims differ from this site's timeline (disclosure on 21 October; 20 October is the leak) and are listed on

the article page for the desk. DBpedia is built from periodic dumps and had no entry for the article; it lags by

months and is not fetched.

For the researcher and the collectors

Every reference of the article is a page the newsroom should hold: when captured, Wikipedia's quotation of it

becomes a quotation this site can make, with the capture's hash. At revision 1381798757: 43 references, 42 with an

address, 16 captured, 26 not, among them W3Techs (the origin of the "70% of websites" figure, contradiction K-001),

the Cyber Security Agency of Singapore, cPanel and CloudLinux, the New York Times, the Wall Street Journal, the

Guardian, the Atlantic, Reuters, NBC. The list is on /wikipedia/php2shell/references/ and in the ontologist's pass

(admin/ontologist/pass.py), which names the uncaptured ones on every run. Ask: add Wikipedia's reference list to

the source watch, so a reference added to the article is captured within the hour.

For the fact-check desk

The statements the article makes with no reference (4 of 39 at the revision read) and the dated Wikidata claims are

on the article page. The attributed statements name who the article says said what; where that party is one this

site has captured, the desk can check the attribution against the capture.

Licence

Wikipedia's text is CC BY-SA 4.0. It is republished here in full, revision by revision, because a delta view needs

it, with the revision id as attribution; the pages and data of the section that carry it are under the same licence,

and say so. The site's rule that no third-party page is reproduced was written for pages that are not licensed for

reuse; this is the one place it is set aside, at the editor's direction, and the editor decides whether it stays.

Added the same day: history, the talk page, a reviewer's reading, and a page for Wikipedia's editors

The editor's direction, later on 25 October: capture the page's history often ("hourly or every 4 hours"), draw its

evolution, read the talk page "since those wikipedia authors are very good", find a way to collaborate with them

without this site editing the article, review the article as a Wikipedia reviewer would against our data, track

Wikidata and DBpedia, and write an admin-site story on adopting their principles. Built:

• The card wakes four times a day (06:05, 10:05, 14:05, 18:05 London, into its own session) and fetches every page in

data/wikipedia/index.json first; the delta is every revision, kept.

• /wikipedia/php2shell/evolution/: every statement as a lifeline; size, references, statements over time; the

Wikidata item's own history. /wikipedia/php2shell/talk/: the talk page as threads, posts, edit requests.

• /wikipedia/php2shell/review/ from data/wikipedia/php2shell/review.json: 15 findings, each a statement, the

policy, the captures, the desk's grade and a proposed change in the edit-request shape; three talk-page questions

answered from the captures. In the graph as finding nodes (questions, rests_on, answers).

• /wikipedia/for-editors/: the collaboration proposal. The rule that decides its shape is Wikipedia's guideline on

large language models (March 2026): no article text, talk-page comment or edit request written by a model is

accepted, and {{edit COI}} requests found to be model-written are declined. So: no agent edits or posts; the

human editor may post, in his own words, disclosed, in the edit-request form, citing independent sources only; this

site asks for no citation or link (WP:SPS, WP:ADV); captures are for checking, the Wayback Machine is the archive

to cite; what we take we attribute (the REUSE wording is on each page); corrections here first.

• data/wikidata-links.json: every party looked up on Wikidata, 25 confirmed by hand; the index page's

"Wikidata and DBpedia" section tracks the item, its history and DBpedia's status.

• The story what-wikipedia-knows-about-being-right on the admin site: eleven principles (BURDEN, directly supports,

SAYWHEREYOUREADIT, RSBREAKING, the ten-year test, words to watch, NPOV and DUE, edit summaries, BRD and 3RR, the

disclosed edit request, the GA checklist), each as it would read in AGENTS.md; and what not to adopt.

Not built

A second article (candidates: PHP, the list of security incidents, the pages of the parties named); the comparison

of correction rates across publications; sentence-level reading of other sources (board 008); the article's other

languages; the /fractal/ navigator and <sg-graph> brought together (board 181).