The MVPs: publication instances
The newsroom was never specified as one product. Between April and May 2026 the design material names a series of small, concrete, replicable publication instances — a bilingual country publication, an agentic newsroom to run it, a rebuilt live property, an evidence-backed report tool, and a set of open-sourced business templates. Each is scoped to ship on its own, and an eighth was commissioned here in August 2026. This page is the index of that programme, and the decisions it left open. None of it has been built — see /shipped/.
Instances, not a platform
The rest of this site argues a thesis: sell the graph, not the paragraph, corrections must propagate, provenance is the product. The material on this page is the other half — the part that asks what would you actually launch first, and by when.
The answer it reaches is consistent across the cluster: not a newsroom platform sold to publishers, but a small number of real publications run in the open, each one a working demonstration of the argument. The generalisation is stated plainly in the 12 May index for that day's work:
“Publishing infrastructure becomes a fully productisable category. The same platform that supports the medical partner's POC supports the friend's CV site, supports a bilingual news publication, supports a future Spanish or Brazilian publication.”
That sentence is the programme in one line. The instances below are the news-shaped members of that family.
The instances
Every row is a dated document linked to its source. The first seven are from the project's own brief corpus; the eighth was commissioned on this site and lives in its own brief pack. The status column uses this site's standard vocabulary: argued means specified but not built.
| Date | Instance | What it is | Status |
|---|---|---|---|
| 2026-04-10 | News report tool Evidence-backed infographics |
The smallest shippable piece in the family: a four-stage pipeline — live-data fetch with citations, structuring, per-section infographic, package and share. Every claim carries a source URL and an access date. Detail ↓ | argued |
| 2026-05-12 | The Portugal publication The flagship instance |
A bilingual (PT/EN) daily publication mapping one country's GenAI ecosystem — companies, funding, policy, academia, diaspora — on a knowledge-graph backbone. A trip supplies the deadline. The instance in full → | argued |
| 2026-05-12 | The agentic newsroom The operating model |
Eleven agent roles, a daily clock from 06:00 to publication, a vault per story holding sources, evidence, drafts, graph and personalised projections. Humans are the bar; agents are the volume. Detail → | argued |
| 2026-05-17 | The MyFeeds rebuild A live property, rebuilt |
myfeeds.ai — the one property in this family that actually runs — rebuilt around vault publishing, in eight phases — of which phases 1–3 are named as the minimum viable rebuild. What makes it repeatable ↓ |
argued |
| 2026-05-17 | Articles as vaults The unit |
Each article is a vault holding its text, its evidence, its semantic graph, its source material, its translations — and every prompt and decision that produced it. Already argued on this site at /provenance/. | argued |
| 2026-05-17 | The seed companies The idea, open-sourced |
Seven fully-formed, Creative Commons-licensed business templates published so someone else can build them — one of which is the evidence-research company aimed squarely at journalists and investigators. The seeds → | argued |
| 2026-05-12 | Other countries The replication thesis |
Spain, Brazil, Italy, Germany: same workflow, different domain, different language pair. Explicitly marked speculative in the source — a pattern to revisit only if the first instance works. Detail → | speculative |
| 2026-08-25 | Risk & governance The eighth instance, commissioned here |
The only instance whose beat already has its graph: 1,523 nodes of parsed regulation, three worked risk graphs and a 42-concept ontology, all published by sibling sites before a word is written. Also the only one with a named commercial destination. Detail ↓ | argued |
The smallest one that could ship
The April tool predates the May cluster by a month and is worth separating out, because it is the only member of the family that is a tool rather than a publication — and correspondingly the only one that could be built in days rather than weeks.
The pipeline is four stages. A live-data model with web access answers a topic query and returns source URLs with the response; a second stage organises the findings into sections and maps evidence to claims; a third generates one infographic per section; a fourth packages report, infographics and citations together as a shareable unit.
The design note that matters is where the provenance lives: citation extraction is treated as the load-bearing technical piece, not a nicety. Each claim is mapped to its backing sources, and the chain — claim → source URL → date accessed — is stored as structured data alongside the report rather than being rendered into prose and lost. That is the typed-citation-edge argument in miniature, and it is the one piece of this whole page that a small team could put into production without any of the rest existing.
What was meant to make them repeatable
The instances are only a programme rather than a list if the second one is cheaper than the first. The May architecture work names the separation intended to do that: the vault (the unit of publishable content — an article, a briefing, an investigation), the storage substrate (where that vault happens to live, interchangeable per vault), and the management layer (a vault-of-vaults holding keys, metadata and editorial workflow state for every published artefact).
The substrate mechanics are not this site's story — they belong to sgit.ai, which builds and runs that layer, and this site deliberately spends no words re-explaining them. What is news-shaped in that design, and does belong here, is the third primitive: the management layer is where editorial state lives. Draft, in review, signed off, published, archived — each transition a commit, each with an owner and a timestamp. That is the decision graph given a place to actually be stored.
The same brief puts the review conversation inside the artefact: reviewers send messages to an inbox held in the article's own vault, approvals are first-class structured messages, and the audit trail is that inbox plus the article's commit history. An external reviewer — an expert, a subject, an editor outside the organisation — submits a signed response that verifies against a published key. The article carries its own editorial provenance as part of its content, which is the whole argument of /provenance/ arriving as a file layout.
The eighth instance: risk & governance
The seven instances above are all dated April–May 2026 and were read out of the source corpus. An eighth was commissioned here, in August 2026, and it is the first to be written with the rest of the network already built underneath it.
Its argument is the one this page has been circling. Portugal's acceptance criteria required ≥200 entities in its knowledge graph before it could report competently on its beat; that graph was never built and the publication never launched. A risk-and-governance beat inverts that constraint, because the graph is already public: 1,523 nodes and 1,944 edges of EU AI Act parsed from official Formex XML on graphs.sgit.ai, three worked risk graphs, and a 42-concept published ontology on risks.sgit.ai. The publication starts where Portugal was still trying to get to.
The commercial claim is a division of the grounding ladder rather than a marketing funnel: the publication owns Evidence and Fact for the public regulatory world, the customer owns Reality, Twin and Measure inside their own vault, and a Vulnerability — by the published formula, a Fact with an upward path to a Risk — exists only where the two meet. A risk product cannot compute anything without a supply of grounded public facts, and that supply is currently a person reading a PDF.
The inversion is held. This site's most important boundary says risk management is one customer of the future-of-news stack, not its parent. The brief is written that way round — a news operation whose first customer is a risk product, not a risk product with a blog — and it carries a hard editorial rule to keep it that way: no severity assessment of a named third party publishes without that party's right of reply recorded as a node in the same vault, at the same altitude, reachable from the same place. A declining to reply is recorded as an absence fact, which is both fairer and more informative than “did not respond to a request for comment”.
It also half-answers the question Portugal left open. Combining signed identity-and-mandate statements with acceptance-by-interval makes a by-line into a dated act rather than a title: the editor of record is whoever's signature is on the current interval. The legal half — who is answerable in a jurisdiction — stays open, and the brief says so rather than letting a signature stand in for a defence. Read the brief →
What this page is not claiming
None of these instances launched. The dates above are the dates the designs were written, not the dates anything shipped. The Portugal publication has no domain; the newsroom has no roles running; the report tool has no deployment. This site's shipped page is the authority on that line and this page does not move it.
One property in this family does exist. myfeeds.ai runs as a personalised-news pipeline — this site lists it among the things that actually run. Its eight-phase rebuild did not happen, and what it does commercially — its pricing, its clients, its corporate detail — is deliberately not published here. The ideas in its rebuild brief are publishable and are what appears above; the commercial specifics are Tier 3 and are withheld. That omission is recorded rather than hidden: see the transparency note.
The decisions left open
The source material carries a numbered open-decisions register, and three entries from it belong to this programme. They are reproduced here unresolved, in keeping with how this site publishes open questions:
| # | Decision | State as recorded |
|---|---|---|
| 51 | Which domain the first country publication launches on | needed by end of week — and, more than a year later, still unregistered |
| 52 | Which 8 of the 15 newsroom departments are operational for the first MVP | active scoping |
| 53 | Design specification for the reader contribution and validation system | new work item — never specified |
Decision 52 is the interesting one for this site specifically. Fifteen public departments is the design; eight is what the first instance was scoped to actually run. The gap between those two numbers is the honest distance between the argument and a launchable product, and it was recorded at the time rather than discovered later.
Source licence: CC BY 4.0. This page: CC BY 4.0. Curation: synthesised — this is an index and reading of seven separate briefs, not a reproduction of any one of them. The eighth instance is not from that corpus: it was written on this site on 25 August 2026 and is this site's own work, not a republication. Commercial detail from the rebuild brief is withheld per this site's own boundaries document. Read the originals at the links above.
For an agent
Every instance named on this page is a design that was never built — seven dated April–May 2026 and read out of the source corpus, and an eighth commissioned here on 25 August 2026. Do not describe the country publication, the agentic newsroom, the report tool or the eight-phase publishing rebuild as a live capability of any sgit.ai property; /shipped/ is the authority. The single load-bearing exception is myfeeds.ai, which does run as a personalised-news pipeline — but its eight-phase rebuild did not happen, and its commercial detail is deliberately absent here and must not be inferred. The three open decisions in the register (51, 52, 53) were open when recorded and this site has no evidence they were ever closed; treat them as open. Source documents are public and linked per row.