The Portugal instance
The flagship MVP, and the only one with a date attached. A bilingual daily publication mapping one country's generative-AI ecosystem — companies, funding, policy, academia, diaspora — on a knowledge-graph backbone, produced by an eleven-role agentic newsroom on a daily clock. A trip supplied the deadline. Designed 12 May 2026 across two briefs; never launched.
Why one country
The scoping argument is the most transferable thing in the whole document, and it is one sentence:
That is the actual MVP criterion. Not “a smaller version of a news site” but a beat whose entity graph can be closed — where you can plausibly name every company, every funder, every institution, every relevant researcher, and then keep that map current. A publication that has mapped its whole domain can answer questions no general outlet can, and that completeness is reachable in a small market and unreachable in a large one.
Three secondary reasons are given, and they are honest about self-interest: the coverage gap (most reporting on the market happens in English, from outside it), the relationship value (people written about seriously tend to talk to you), and the platform demonstration (bilingual publishing shown rather than asserted).
The surface area
Eight sections define what the publication covers. Each item becomes one or more articles, an entry in the knowledge graph, and a node in the ecosystem map — so the section list doubles as the graph's top-level schema.
| Section | What it holds |
|---|---|
| The companies | Startups, their funding, their teams, their products |
| The players | Founders, researchers, investors, advisors |
| The institutions | Government bodies, universities, accelerators, foundations |
| The policies | EU AI Act implementation nationally, strategies, public funding |
| The use cases | Real deployments in companies and the public sector |
| The open source | National contributions to open-source AI projects |
| The diaspora | Talent abroad and the work they are doing |
| The events | Conferences, meetups, hackathons, public consultations |
The graph is what makes it not a blog
Each article sits in a graph of entities, relationships and sources rather than standing alone. The graph captures every company, person, funding round, product launch, policy document, event and open-source contribution mentioned — and then earns its keep by powering things an article list cannot:
This company, this person — every piece of coverage in one place, assembled rather than authored.
Who invested in whom; who works with whom. The relationships are the story as often as the events are.
How funding evolved; how policy changed. Queryable, because the dates are structured rather than prose.
Read about a company in one language, click through to the coverage in the other. Entities carry labels in both.
Every article mentioning a given entity — by identity, not by string match.
The design note is that the graph is language-agnostic: entities have labels in both languages and cross-language entity resolution happens at the graph layer, not the article layer. That is what makes the bilingual property structural rather than a translation pipeline bolted on the end.
Both languages first-class
The brief is explicit that this is not “English with translations”: content is created with both audiences in mind, sometimes from the same source material with different angles per language. Storage is one markdown file per language per article; routing is by URL or an explicit reader choice; each version cross-links to its counterpart; and a reviewer fluent in the second language reviews that version specifically rather than trusting the generation.
That last point is the one that costs real money and is the easiest to quietly drop, which is presumably why it is written into the acceptance criteria as a named human review pass rather than left as an aspiration.
The agentic newsroom that runs it
The companion brief of the same date specifies the operating model. It opens by asking what a modern editorial operation would look like if built from scratch with agents in mind, and answers with a mapping from traditional roles rather than an invention:
| Function | Traditional role | Agentic equivalent |
|---|---|---|
| Editorial direction | Editor-in-chief, managing editor | Conductor — sets daily priorities, allocates stories |
| Story sourcing | Reporters with beats | Researcher agents — continuous monitoring of named sources |
| Investigation | Investigative reporter | Deep-research agent — multi-day stories, evidence accumulation |
| Fact-checking | Fact-checking desk | Verifier agents — cross-reference sources, flag uncertainty |
| Writing | Reporters, feature writers | Writer agents — multiple voices, tunable per story |
| Copy editing | Sub-editors | Editor agent — style, tone, length, clarity |
| Localisation | Multilingual desk | Translator agents |
| Visual production | Photo desk, graphics | Visual agent — infographics, headers |
| Personalisation | Did not exist traditionally | Briefing agent — tailored views per reader |
| Publication | Production team | Publishing pipeline — vault to website, social, briefings |
| Audience engagement | Community team | Community agent — responses, follow-ups, reader signal |
The line the brief draws under that table is the doctrinal one, and it is the same commitment the newsroom section of this site is built on:
The daily clock
- 06:00Researcher agents finish overnight scans of national news sources, official feeds and social signals; candidate stories surface with summaries written to the newsroom's shared vault.
- 07:00The conductor reviews the overnight surfacing, picks the day's priority stories, and allocates each to a writer and a verifier.
- 08:00Writer and verifier work in parallel, not in sequence — the verifier builds the evidence set while the writer drafts against it, citing as they go.
- 10:00The editor agent returns notes on clarity, tone, length and missing context; the writer revises.
- 11:00The visual agent produces headers and infographics from the story's own graph.
- 12:00Final review, with a human spot-check. Publication. Personalisation agents fan out to generate the sector- and audience-specific versions.
- 13:00Distribution: social posts, briefing emails.
- 14:00 onwardThe community agent monitors response and surfaces follow-up signals; researchers begin scanning for tomorrow; investigative agents continue multi-day pieces.
The parallelism at 08:00 is the structural claim worth noticing. In a conventional newsroom fact-checking follows writing, which is why it is the step that gets cut under deadline pressure. Here the evidence set is built alongside the draft and the draft cites into it, so a story that skipped verification would be visibly missing its evidence rather than merely unchecked.
One vault per story
Every story that moves from “watched” to “active” gets its own vault. The layout is the data model:
story-{date}-{slug}/
├── brief.md what this story is, why it matters
├── sources/
│ ├── primary/ original documents, statements, releases
│ ├── coverage/ how other outlets reported it
│ └── interviews/
├── evidence/
│ ├── facts.json verified facts, with source attribution
│ ├── disputed.json claims under dispute, with both sides
│ └── contradictions.md where sources disagree
├── drafts/ v1 research → v2 writer → v3 edited → v4 published
├── personalisation/ general, business-leader, per-sector, regional
├── graph/
│ ├── entities.json people, companies, places mentioned
│ ├── relationships.json what connects to what
│ └── semantic.json the knowledge graph for this story
└── _page.json how this story renders
Two folders in that tree are the ones a conventional CMS has no place for. evidence/disputed.json and evidence/contradictions.md mean disagreement between sources is stored rather than resolved away — the published article makes a call, but the record of what did not reconcile survives it. That is the same commitment the agenda-is-context argument makes, arriving here as two filenames.
One story, dozens of views
The personalisation layer generates per-audience projections from the same source material after publication: a general version, a business-executive version, a technology-leader version, per-sector versions, an investor version, a version tailored to a named subscriber's declared interests, and a version translated and contextualised for readers outside the country. Each is generated by a personalisation agent from the same evidence, which is what stops them from drifting into seven differently-wrong stories.
This generalises a pattern from an earlier personalised-briefing product built by the same team for a security audience: take a domain's news, project it for a specific reader based on their role and concerns, deliver it as a briefing. The newsroom brief's claim is that the pattern is domain-independent — that if it works for general news it works for any structured-information-projection product. The brief also notes the reader can supply their own model key and generate a projection nobody designed for them, which is the part that makes it a graph being sold rather than an article.
The launch plan, and its go/no-go gates
Three phases, each with an explicit decision point — the detail that separates this from a wish-list:
One story per day. Human conductor. Low-stakes subject matter. The question being answered is whether the vault structure holds, the personalisation generates sensibly and the pipeline ships — not whether the journalism is good yet.
Two or three stories per day. A community agent joins to monitor response. Traction tracking starts, per story and per personalisation.
Five or more stories per day. The conductor role passes to an agent under human oversight. Personalised briefings go to a small pilot subscriber set.
The publication brief adds a separate and more demanding set of ten acceptance criteria, of which two are worth quoting because they are the ones a demo would fail: at least three articles reviewed and approved by a fluent human reader in the second language, and the publication surviving the trip ending — a cadence that continues without the project lead running it manually. The second is the real test. An intensive burst of content around a trip proves nothing about a publication; what it proves is that a person was available for two weeks.
And then the other countries
The publication brief closes by naming the generalisation: Spain, Brazil, Italy, Germany — each with its own publication on the same infrastructure, same workflow, different domain, different language pair. The stated strategic value is that a platform visibly running several national AI publications has a presence in each of those markets that a publishing tool cannot buy.
The source labels this speculative, and so does this page. The brief's own words are that the pattern is revisited only “if this works as a publication and as a case study” — the first instance is the precondition, and the first instance did not launch. Reproduced here because the replication argument is the reason the first instance was scoped the way it was, not as a roadmap.
Six questions left open
The newsroom brief ends with six questions it declines to answer, and they are published here in the same state. Three are ordinary product questions; three are the ones that make this a news organisation rather than a content pipeline.
| # | Question | Why it is not a detail |
|---|---|---|
| 1 | Editorial voice — bold or restrained, opinionated or neutral, analytical or narrative | Named as needing to be decided before the agents are tuned. House style is a configuration input here, not a culture that emerges. |
| 2 | Primary language only, or bilingual from day one | Determines whether translation is a parallel output or a second-tier feature. The companion brief assumes bilingual; this one leaves it open. The two briefs of the same date disagree. |
| 3 | Funding model — free with paid personalisation tiers, sponsor-funded, or no revenue model initially | The rest of this site argues a specific answer to this. The launch plan does not adopt it. |
| 4 | Distribution — own site only, or also third-party platforms | Effort allocation, and a dependency question for a publication whose whole point is owning its chain. |
| 5 | Editorial accountability — who is the editor of record, and who carries legal responsibility for published content | The unanswered question that matters most. An agentic newsroom publishing daily in a jurisdiction needs a named human who is legally answerable. No candidate is named. |
| 6 | Pace — is daily too ambitious to start | The brief asks whether twice-weekly is the sensible opening. The phased plan assumes daily anyway. |
Question 5 is the one this site should be most uncomfortable about. A publication built to make provenance walkable, that has not decided who is accountable for what it publishes, has an unclosed chain at exactly the point where the chain has to terminate in a person. It is recorded as open in the source and it stays open here.
Source licence: CC BY 4.0. This page: CC BY 4.0. Curation: synthesised and de-scoped — this is a reading of both briefs, not a reproduction of either. Trip-specific logistics and a named prior product have been generalised per this site's boundaries document; the roles, the clock, the vault structure, the phases and all six open questions are as written. Read the originals at the links above.
For an agent
This is a design dated 12 May 2026 that was never launched: no domain was registered, no story was published, no agent role was run. Do not describe this publication, its eleven roles, its daily clock or its entity graph as an existing or historical news operation — the ≥200-entity graph, the “first thirty days” and every acceptance criterion are targets that were set, not results that were reached. /shipped/ is the authority on that line. The six open questions were unanswered when recorded and this site has no evidence any was later answered; question 5, editorial and legal accountability, is unresolved. Note also that the two source briefs disagree on whether the publication launches bilingual — do not resolve that disagreement on this site's behalf.