newsroom.sgit.ai — The Future of News

News is failing not because there is too little information but because there is no walkable chain from a claim to its evidence, no way for a correction to reach what it disproved, and no way to pay the person who did the original work. A story is a graph that accumulates evidence, perspectives and confidence; every article, translation and infographic is a projection of it. Sell the graph, not the paragraph.

Site version: v0.3.13 (13 September 2026). Published by the sgit project, which is building the stack this site argues for — participant disclosure at /about/participant.html. Site content CC BY 4.0; the republished library material is CC0 1.0 Universal at source (docs.diniscruz.ai) — state the source licence per page.

The honest sentence

Most of this site is an argument, not a product. THREE THINGS RUN, and each says its own limits on its face: /portugal/ (Portugal Startups — sources fetched, frozen and hashed; a graph; three stories; human editor of record; beta, no legal review), /databases/ (SQLite and a SPARQL store in the reader's browser over /portugal/'s files; no server, no store of record) and /governance/ (the Governance Wire — graph, vocabulary, source register and gate run; its ingestion and verification layer is NOT built, so no claim there is a checked reading of any regulation; no human review). The articles in /library/ are real and dated; the rest of the newsroom is a design. /shipped/ states the line between them in full. Do not describe Trust-as-a-Service, the decision graph, per-story cost ledgers, contextual-validation billing, or any newsroom department as a live capability of any sgit.ai property. The same applies to everything in /mvps/: the country publication, its eleven agent roles, the evidence-backed report tool and the seven seed companies are designs dated April–May 2026 that were never launched — their targets ("first thirty days", a ≥200-entity graph) are goals that were set, not results that were reached. Of the three that run, only /portugal/ has a human editor of record; /governance/ has none; none has had legal review. The nav names the three as their own top-level groups.

Properties agents may rely on

• Every page deriving from previously published material carries a visible provenance block: original publication date, original link, original authors (including any AI co-authorship), and an honest curation label (verbatim / edited / excerpted / synthesised). Enforced by the pre-release gate, not just documented.
• Ten core articles, 68,846 words, are already public and dated at docs.diniscruz.ai (Feb–Oct 2025), a year before this site's design material was written — see /library/. Several republished sections on this site are this site's own SYNTHESIS of that material, not verbatim reproduction; fetch the docs.diniscruz.ai canonical URL directly for complete original text.
• The payment rails described in /economics/rails/ (x402, Cloudflare Monetization Gateway, AWS Bedrock AgentCore Payments) are real, operating third-party infrastructure, independently verifiable — unlike almost everything else this site describes.
• Seven open questions and six honest tensions are published unresolved at /network/#open and /network/#tensions — check there before assuming any of them has been answered.
• Every page ends with a pasteable "for an agent" block, except /admin/ pages and /about/participant.html.
• The three running instances publish machine surfaces you can fetch without rendering a page: /portugal/data/.json and /portugal/data/triples.nt; /databases/data/.json; /governance/data/*.json. The SQL and SPARQL consoles publish window.__tools after a tool:ready event, read-only, so an agent driving a browser can query rather than scrape.
• /llms-full.txt is this file plus the front page plus every markdown document in the brief pack, concatenated — one fetch for the whole document set, for when link-following is unreliable. The two structured-data files are deliberately left out of it and should be fetched directly.

Site map

• /index.html — front page: the claim, the 10,000-hours story, the proof strip
• /thesis/index.html — the thesis: sell the graph, evidence not truth, the author is the oracle
• /corrections/index.html — corrections must propagate (the site's most distinctive argument)
  • /corrections/the-claim-that-would-not-die.html — the 10,000-hours case in full
  • /corrections/242-papers.html — the same failure, measured: 242 papers, >220,000 citation paths
  • /corrections/how-a-graph-answers-it.html — supersede-never-delete, typed citation edges
  • /corrections/agenda-is-context.html — the self-critique: the graph has an agenda too
  • /corrections/staleness-in-the-wild.html — the EU AI Act's "consolidated" text, dated 2024
• /provenance/index.html — provenance is the product
  • /provenance/show-your-work.html — fifteen public newsroom departments
  • /provenance/a-worked-story.html — one story, itemised: £8.40, 6h 23m
  • /provenance/the-decision-graph.html — named ownership at every editorial step
  • /provenance/articles-as-vaults.html — each article as text + evidence + graph + every prompt
• /library/index.html — the published record, 2025 → present, chronological
• /economics/index.html — paying the fact creator
  • /economics/paying-the-fact-creator.html — the 60/25/10/5 split; contextual validation
  • /economics/trust-as-a-service.html — the fact-certifier as a payable, warranted role
  • /economics/rails.html — x402, ~200ms settlement, zero protocol fees (real, operating)
  • /economics/micro-and-nano-payments.html — the 2025 argument, synthesised, paired with the rail
• /rights/index.html — content rights: CC-Signed, the licence family with a legal stick
• /newsroom/index.html — roles, the daily clock, fifteen departments, the craft doctrine
• /mvps/index.html — the publication instances: the newsroom was specified as a series of small launchable publications, not one platform. Seven instances, April–May 2026, none built
  • /mvps/portugal.html — the flagship: a bilingual country publication run by eleven agent roles on a daily clock. Its two source briefs disagree on whether it launches bilingual, and its open question 5 — who is the editor of record and legally accountable — was never answered
  • /mvps/seed-companies.html — seven CC-licensed business templates published so others can build them; seed 4 is an evidence-research company for journalists and investigators
• /portugal/index.html — PORTUGAL STARTUPS (beta): a publication mapping the Portuguese startup ecosystem, first beat Startup Summit Lisbon 2026 (17-18 September 2026, Beato Innovation District). THE IMPORTANT DIFFERENCE FROM /governance/: sources here are PRIMARY and ANCHORED. Every page of the source site is fetched, frozen to a dated snapshot in the repository and hashed with SHA-256, and claims are read from the frozen copy rather than the live network. This is the fetch-freeze-hash path /governance/ specifies and cannot run. It also has a NAMED HUMAN EDITOR OF RECORD (Dinis Cruz) who reviews every page before it publishes, because this beat is about named people and named companies.
  • /portugal/graph.html — THE GRAPH: 329 nodes / 1,106 edges, 16 types, 19 verbs — the event, 64 speakers, 61 derived organisations, 16 sessions on 3 stages, 88 frozen sources across 2 snapshots (64 of them the speakers' own pages), 7 pieces of press coverage, our stories, 23 Topic nodes (the event's own tags) and 50 derived tag nodes (Industry, Technology, Idea, Service, Product — by lexicon, every edge carrying its matched words). Every edge is a verb with a named inverse IN ENGLISH AND PORTUGUESE (ontology.json), so a path reads as a sentence either way. Nodes arrive in PACKS the reader switches on block by block. The page publishes window.__graph (read + view methods, no writes) after a tool:ready event, the same convention as graphs.sgit.ai's universe reader. Fetch /portugal/data/graph.json and /portugal/data/ontology.json directly
  • /portugal/explorer.html — THE FILES: every file the section is built from, from /portugal/data/manifest.json with size and SHA-256, rendered / as a graph / raw. Frozen third-party pages are listed and hashed but never rendered
  • /portugal/connections.html — CONNECTIONS: who should be talking to whom, who could be buying from whom and who could help whom, as THREE ORGANISATION-LEVEL SQL QUERIES over two things read from each speaker's own frozen page — the event's own Topics list (verbatim, 23 distinct) and the words matching a PUBLISHED LEXICON of 57 patterns (industries, technologies, ideas, services, products; the matched words only, never the sentence). Rows at build in /portugal/data/connections.json with the SQL beside them; the same SQL runs unchanged in /databases/sql.html. NOTHING HERE IS A CLAIM ABOUT A PERSON: the unit is the organisation, and the notice refuses characterisation of named individuals. Gate 17 re-runs every pattern on the frozen bytes each build; gate 16 checks every topic is on the page verbatim
  • /portugal/summit/index.html — the event as it describes itself. Distinguish TARGETS from COUNTS: the event's own wording is "targeting 2,000+" attendees and "150+ speakers planned", while its published list names 64. Do not report a target as a result.
  • /portugal/summit/people.html — all 64 published speakers with role, organisation, the event's own Topics for them (verbatim) and a link to each speaker's own page. Biographies are deliberately NOT reproduced; the lexicon reads them and keeps only matched words
  • /portugal/summit/orgs.html — 61 organisations DERIVED from the speaker cards; 2 are placeholders ("Independent") flagged rather than removed
  • /portugal/summit/changes.html — the same pages frozen twice and diffed: 5 speakers added and 1 removed between 8 and 13 September. REMOVALS CARRY NO REASON and none may be inferred
  • /portugal/sources.html — the register: every frozen file with its SHA-256, byte count and retrieval time, and how to check any claim against it. Two kinds of source: the event's own pages, and third-party pages ABOUT the event — none a registry or funding dataset, so nothing here describes the ecosystem beyond this one event
  • /portugal/checks.html — what re-reading established, including two different sets of stage names published on the same site, and the claims that cannot be verified from the source
  • /portugal/method.html — fetch, freeze, hash, extract, diff
  • /portugal/team.html — seven agent roles plus the named editor of record
  • /portugal/notice.html — THE DATA-PROTECTION NOTICE. This section names 64 people it did not get the data from. Lawful basis is LEGITIMATE INTERESTS and the journalistic derogation is EXPRESSLY NOT CLAIMED: Article 24(3) of Portugal's Lei 58/2019 conditions it on national rules on access to and exercise of the profession, which this publication does not meet. Personal contact details of any kind are refused at EXTRACTION time and gate check 11 fails the build if one appears in the data. Removal on request is unconditional and no reason is asked for. Carries a Portuguese-language summary. Not legal advice
  • /portugal/about.html — beta, one beat, two kinds of source, no legal review, nothing in Portuguese
  • /portugal/stories/a-record-attempt-the-agenda-does-not-mention.html — two Portuguese outlets report a 48-hour Guinness pitch marathon on 16-17 September; the event's own agenda, four days out, does not list it. Both frozen; neither wrong yet
  • /portugal/stories/five-names-arrived-and-one-left.html — the speaker list went 60 → 64 between 8 and 13 September; both copies held and hashed; the removal carries no reason
  • /portugal/stories/two-names-for-three-stages.html — the agenda and the AI-summary page name the same stages differently, on the same site, on the same day
  • /portugal/team/<role>/index.html — one page per role (researcher, extractor, diffwatch, cartographer, writer, editor, publisher): definition, doors, refusals
  • /documents/pt-newsroom.html — the COMMISSIONING BRIEF for pt.newsroom.sgit.ai, addressed to the agent that will build it; section 16 records the home-page direction chosen on 13 September; section 17 points at THE BRIEFING PACK, /briefs/pt-newsroom-pack.zip (unpacked under /briefs/pt-newsroom-pack/): the brief, the design sources and fonts, the code to inherit, the newsroom operating model, CLAUDE.md and settings for the new repository, the prompts and repository skills for the scheduled run and the editor's review, the three ways to schedule it with a cron workflow, the acceptance test, and a handover. Start at 00__READ-ME-FIRST.md
  • /portugal/data/{graph,ontology,manifest,coverage,sessions,event,people,orgs,sources,changes,checks,team,stories,notice,topics,lexicon,connections}.json and /portugal/data/triples.nt — machine surfaces
  • /portugal/sources/frozen/<date>/<page>.snapshot — the frozen byte copies themselves. Stored with a .snapshot extension because they are EVIDENCE, not pages of this site: unmodified bytes of another organisation's pages, held so a claim can be verified, never republished as browsable content
• /pt-newsroom/index.html — THE HOME PAGE OF pt.newsroom.sgit.ai, AS DESIGNED: the site does not exist yet; this is its chosen front page (13 September 2026) rendered at 1440px and 390px from the design sources /pt-newsroom/design/{Main,MainPhone}.dc. A classic broadsheet, natively Portuguese, drawn with /portugal/'s real data of that day and NOT maintained — take the structure, not the numbers. The design system is section 16 of /documents/pt-newsroom.html. Nothing on it is a claim.
  • /pt-newsroom/directions.html — THE FOUR DIRECTIONS, ARCHIVED: all four home pages drawn that day (broadsheet — chosen; dense ledger; dark graph-first; magazine with claims underlined), each rendered from its source in /pt-newsroom/design/ with the motivation and trade-off written on the canvas. Three were NOT chosen and are not to be built from; they are kept so the choice can be re-read. Seven typefaces vendored under /assets/fonts/, all SIL OFL
• /databases/index.html — DATABASES WITH NO SERVER (beta): two real database engines running in the READER'S BROWSER, compiled to WebAssembly, over the JSON files /portugal/ is built from. No server, no upload, no store of record: the files are the database, the engines are readers, the store is discarded when the tab closes. The pattern is the estate's (sgit.ai RiskMandate: "the browser becomes the database"; graphs.sgit.ai Regulation Graph: SQLite over WASM, client-side and ephemeral) and graphs.sgit.ai's "NOT A GRAPH DATABASE PITCH" is inherited: JSON stays the source of truth. Both consoles publish window.__tools.{sql,sparql} after a tool:ready event, read-only, and fetch only same-origin files.
  • /databases/sql.html — SQLite (sql.js 1.14.2, MIT, vendored): 15 tables built on load from /databases/data/tables.json (the loader spec, which names the file and field behind every column); 16 worked queries from a count-by-class to a recursive two-hop traversal and the connections formula, each with the row count it returned AT BUILD (/databases/data/queries-sql.json). The build executed every one through Python's sqlite3 against the same spec; a failing or empty query fails the build
  • /databases/graph.html — Oxigraph (0.5.11 web build, MIT OR Apache-2.0, vendored): SPARQL 1.1 over /portugal/data/triples.nt (N-Triples, IRIs under https://newsroom.sgit.ai/portugal/ ↗ {id,verb,type,prop}/; the ontology is IN THE STORE, inverses declared with owl:inverseOf and walked with ^, labels @en and @pt; 4,001 triples). 13 worked queries incl. a Portuguese path read aloud, a property-path two-hop walk, a CONSTRUCT of the inverse edges and an ASK that no edge is symmetric — each shown beside the same question in Cypher, which is displayed and NOT run (Kùzu's WASM build is 73 MB unpacked; declined and recorded). Validated at build with pyoxigraph
• /governance/index.html — THE GOVERNANCE WIRE (beta): the first MVP of the risk-and-governance publication, self-contained under /governance/ so it can move to its own domain later. It is FULLY AGENTIC — seven agent roles, no human reviews any page before publication — and every fact in it is SECONDARY: taken from another publication's reading of a primary text, with nothing frozen, hashed or anchored. Do not cite it for any regulatory fact; follow the source link on the node and cite the original
  • /governance/graph.html — every node and edge, and the vocabulary they conform to
  • /governance/sources.html — where every fact came from and whether anyone read the original
  • /governance/method.html — what the graph/verification/linking layer adds, and the 13 gates
  • /governance/team.html — seven agent roles, each with a centre of gravity and refusals
  • /governance/team/<role>/index.html — one page per role: the definition in full, the doors only that role can open, and what is on its desk right now. Source: team/<role>/role.md
  • /governance/newsroom/index.html — THE FLOOR: the room itself, as a point-and-click scene. Seven agent desks joined by the pipeline route; each answers from desk.json. Not decoration — the workload it reports is the live one
  • /governance/newsroom/workflow.html — THE STATE MAP: nine states from search result to published page, each with the door it must pass. The frozen state is BLOCKED and nothing has ever passed it, which is why every fact here is secondary and nothing is anchored
  • /governance/research/index.html — the Researcher's runs, published whether or not anything came of them
  • /governance/research/<date>.html — one run: what was searched, what resolved, what was actually READ, and what our own egress blocked. Two Council of the EU pages return 403 to us; that is recorded as a fact about our reach, NOT about those pages
  • /governance/about.html — beta, no legal review, no human review, nothing anchored
  • /governance/stories/<slug>.html — the stories; prose in /governance/content/<slug>.md
  • /governance/data/{graph,ontology,sources,stories,team,workflow,desk,research}.json — the machine surfaces; fetch these rather than parsing the pages
• /shipped/index.html — what runs vs what is argued. Read this before citing anything as live
• /network/index.html — sibling boundaries, the Risk Mandate inversion, open questions
• /documents/index.html — the brief pack, readable in-page; raw markdown at /briefs/
  • /documents/<slug>.html — one reader page per pack document (generated projections; the file under /briefs/ is the source of truth if the two ever disagree)
  • /documents/newsroom-floor.html — a DEBRIEF, not a brief: how and why the agentic team was rendered as a point-and-click room, written for the agents of the other *.sgit.ai sites so they can build the same thing. The transferable rule is that every word the room speaks is derived AT BUILD TIME from the same files the pipeline runs on, and a gate fails the build if the route drawn through the desks stops matching the declared pipeline. Includes the porting recipe, the four defects we shipped into, and what a single implementation on a single day has NOT proven. Raw source: /briefs/10__the-newsroom-floor.md
  • /documents/pt-newsroom.html — a COMMISSIONING BRIEF addressed to the agent that will build pt.newsroom.sgit.ai, a natively Portuguese newsroom. The load-bearing finding: the journalistic derogation is NOT available to an unaccredited agent-produced publication, so the basis is legitimate interests, which costs a balancing test, a published Portuguese notice and a named accountable human. Also cuts fifteen departments to three, settles the subdomain by wildcard-certificate mechanics, and rules that in a natively Portuguese publication the graph EDGE VERBS are Portuguese verbs. Raw: /briefs/11__pt-newsroom-commissioning-brief.md
  • /documents/summit-archive.html — THE CONSOLIDATED ARCHIVE of the Startup Summit Lisbon 2026 beat: 160 files — every page, data file, build script, story, role page, brief and frozen source — each with its live URL, its GitHub URL, its SHA-256 and a line saying what it is, all verified live and byte-identical when cut. Bundle: /briefs/summit-archive.zip (210 files, 1.5 MB), unpacked under /briefs/summit-archive/. Four narrative documents: the chronology (two captures, 8 and 13 September), every published number beside the file it derives from, the method and its eighteen gates, and three things that went wrong. THE BOUNDARY IT STATES: the register stops on 13 September, four days before the doors; this publication holds no capture from during or after the event and makes no claim about it
  • /documents/pt-transfer.html — A SIBLING TRANSFER, addressed to the agents of pt.newsroom.sgit.ai (github.com/SGit-AI/SGit-AI__Website__Newsroom__PT), read at its v0.3.1. What that site built and this one lacks (a read-only JSON API with OpenAPI, the delivery quarantine as a build gate, the accent and Portuguese-path gates) and what goes the other way: two frozen Summit captures from 8 and 13 September (60 then 64 speakers) that turn its lead story — explicitly waiting for a second capture of 70 — into three captures in six days; three press pages plus one excluded WITH ITS REASON; and a story standing on two independent publishers. THE RULE IT PROPOSES: evidence transferred between sibling publications stays evidence only if the provenance travels with it and is published — the bundle at /briefs/pt-transfer.zip carries obtido_por, obtido_em, url and the origin register's SHA-256 per file, all fourteen re-verified before packaging
  • /documents/research-brief-chatgpt.html and /documents/research-brief-perplexity.html — two RESEARCH BRIEFS for outside assistants: what pt.newsroom.sgit.ai needs (eight sections, in priority order), how to search Portuguese sources, the rules on people, and the JSON contract for handing back LEADS WITH PROVENANCE — /briefs/pt-newsroom-pack/08__research-briefs/ research-schema.json (JSON Schema 2020-12) with example-delivery.json. Nothing a delivery contains is cited until the newsroom freezes the page and re-finds the excerpt in the bytes
  • /documents/risk-governance-newsroom.html — the pack's only FORWARD-LOOKING document: a commissioning brief for a risk-and-governance publication built on this network, whose commercial destination is riskmandate.ai. It is a publication to build, not one that exists. Its architectural claim is a split of the risks.sgit.ai grounding ladder: the publication owns Evidence and Fact, the customer owns Reality/Twin/Measure, and a Vulnerability exists only where the two meet
• /about/participant.html — participant disclosure; two years of disclosed model co-authorship
• /admin/index.html — how this site is built
• /admin/comms.html — open requests (N1–N6) and standing tasks (T1–T8), in public
• /admin/versions.html — release history
• /llms-full.txt — the whole document set in one fetch

Read before repeating a fact from this site

Every /library/ entry and every provenance block states the ORIGINAL publication date — never the date it reached this site. A version number or a fact about a sibling *.sgit.ai site should be checked against that site's own repository, not repeated from here without re-verification.