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. Make the editorial process public, make the citation chain typed and signed, and two things follow: corrections propagate, and facts become billable upstream. That is the future-of-news stack: sell the graph, not the paragraph.
Most articles do not provide evidence, they provide a link. The link is never followed, and it could go to a site that no longer exists. There is no reference of what references what.
The 10,000-hours case. A 1993 study of violin students at a Berlin academy found the top group had practised roughly 10,000 hours by age 20 — an average, not a threshold; roughly half the top group had not reached it. The mechanism was deliberate practice, not accumulated time: one chess player reached master level in ~3,000 hours, another needed >20,000. Popularised in 2008, the number became a threshold in public memory. The original researcher called it "catchy rather than meaningful — it could as easily have been eleven thousand," and spent much of his career trying to correct it: books, articles, interviews, an open letter. None of it ever attached to the claim. The person best placed in the world to say that is not what my study showed had no channel, no standing at the point of use, and no economic reason to keep doing it beyond his own conviction.
The full story, and what a graph would have done differently →
In a document, a correction is a new document. Nothing that cited the original knows. In a graph, a correction is an edge — and the question how much of what I believe rests on claims that have since been corrected? becomes answerable. No document set can answer that. A graph answers it as a traversal. This is the inverse of how misinformation works today: currently a false claim propagates virally and the correction barely travels. In this model, the correction propagates with the same force as the original claim.
Not projections. A measured citation network, a worked production ledger, and a payment rail that went live in the eighteen months since the thesis was first published.
The article, the infographic, the translation, the per-sector briefing — all projections of one accumulating story node. The thesis →
The editorial process is public: 15 departments, a per-story cost ledger, a decision graph with named ownership. Provenance →
Superseded claims are marked from a date and never deleted. Staleness is priced, not silent. Corrections →
Not the last-mile publisher. A worked 60/25/10/5 upstream split, and a real payment rail to run it on. Economics →
The fact-certifier: a distinct, payable, warranted role. Weight from independence, not rank. Trust-as-a-Service →
CC-Signed: break the signature chain, break the licence. A legal stick, not just a norm. Rights →
Agents absorb production grunt-work so more humans become affordable, funded per contribution. The newsroom →
Lifting text into concepts is decompilation, not compilation. Disagreement is the product. Meaning →
This site ships most days, and every release is tagged, gated and recorded. The newest six:
The gate inside the published vault was broken again, by the release that added the app.
v0.5.4 · 29 Sep 2026One vault, not two — and it opens as an app. Plus the release news on the front page, and the “design, not built” badge retired.
v0.5.3 · 29 Sep 2026The outreach brief: what the collaborating agent is asked to do, to whom, in what order, and with what words.
v0.5.2 · 29 Sep 2026The two vaults listed and described the way sgit.ai lists its thirty-seven — and the live page was pointing at a vault that did not exist.
v0.5.1 · 29 Sep 2026Cloning our own published vault with our own published key found two defects that every internal check had missed.
v0.5.0 · 29 Sep 2026The fractal layer: all 1,070 sentences of the twenty-one op-eds, typed, anchored and connected — plus a page for each piece, and LinkedIn routes.
Everything above is an argument. These four are running instances of it, each a site inside the site, built by agents. Each says on its face what is real and what is not, and which of them a human reviewed before publication.
A publication mapping the Portuguese startup ecosystem, first beat Startup Summit Lisbon 2026. Every source fetched, frozen and hashed; 88 frozen pages, a graph of 329 nodes with Portuguese verbs, three stories, a data-protection notice with a named editor of record. The wire → The graph → Connections →
The twenty-one op-eds WAN-IFRA commissioned for World News Day and made free to republish — fetched, frozen, hashed, described and graphed here, and republished nowhere. All twenty-one grant permission; none grants it under a licence with a name, and none of the twenty-one uses the schema.org license field its pages already ship. The argument → The corpus → The finding → The aggregate →
SQLite and a SPARQL 1.1 store running in your browser over the Portugal section’s own JSON files, compiled to WebAssembly. The files are the database; the engines are readers. Every worked query ran at build and prints the count it returned. The argument → SQL → SPARQL →
The risk-and-governance newsroom as a running MVP: seven agent roles, a nine-state workflow, research runs published whether or not anything resolved, and a point-and-click floor where each role is a desk you can talk to. The wire → The floor → The state map →
The commissioning brief for a natively Portuguese newsroom mapping Portugal’s AI landscape: three departments, the law it has to satisfy, the first three articles, and the home-page direction chosen on 13 September. Read the brief → The home page, as designed →
This does not launch as a manifesto. Ten core articles — 68,846 words — were published on docs.diniscruz.ai between February and October 2025, a year before the design material behind the rest of this site was written. The micropayments argument is from April 2025, fifteen months before the payment rail that makes it real. The full chronology, with the original date on every entry →
PROPOSED in the project's own reality tracker. The line between the two, stated plainly →newsroom.sgit.ai argues that a story is a graph and an article is one projection of it — and that an agent buying facts by API is the first worked example of its own product, not a hypothetical. Four things run — /portugal/, /world-news-day/, /databases/ and /governance/, each stating its own limits — and the rest is a design: read /shipped/ before citing anything here as a live capability. Every page that republishes or derives from previously published material carries a visible provenance block naming the original date and linking to the original — the site's own demonstration of the argument it makes. This site is published by the sgit project, which is building the stack it argues for: read the participant disclosure before treating any page here as neutral. llms.txt is the whole agent surface.