## Mission
Vroeg serves a specific reader: the Belgian or EU investor, policymaker,
or AI practitioner who, for a decision, needs to know quickly **what
changed this week in EU AI policy or AI markets, how certain it is, and
where the primary sources are**.
Not for the general public. Not for consumer AI news. Not for hype.
## Scope
Vroeg covers only:
- **EU AI Act enforcement** — application, AI Office output, delegated
acts, GPAI codes, Member State implementations, EDPB opinions.
- **Belgian + Dutch AI capacity** — imec, ASML, KU Leuven, UGent, VUB,
TU Delft, INRIA-Lille, regional AI startups.
- **EU AI sovereignty** — semiconductor supply chain, Chips Act,
Mistral, Aleph Alpha, GPU allocations.
- **AI in regulated EU sectors** — healthtech (EMA), fintech (DORA),
defense + dual-use (EDF, NATO DIANA).
- **Comparative AI policy** — EU vs US vs China, evergreen content.
Out of scope: general AI news without a Low Countries angle; consumer AI;
LLM hype cycles; tech-CEO-quote pieces; opinion without primary-source
foundation.
Source ladder
Every claim in a Vroeg article must rest on at least one source from
the following hierarchy, in decreasing preference:
1. **Primary legislative text** — EUR-Lex, Belgian Gazette,
`journal-officiel.gouv.fr`, EU Official Journal, COM documents.
2. **Official authority output** — EU AI Office, EDPB, EBA, EMA,
FSMA, FOD Economie, Dutch AP, BeCert, BeIA.
3. **Listed-issuer filings** — SEC filings, ASML investor relations,
ECB stress-tests, Bank of England Working Papers, EBA risk dashboards.
4. **Peer-reviewed academic work** — arXiv (with preprint disclaimer),
OpenAlex, conference proceedings (NeurIPS, ICML, ACL, COLM, ICLR).
5. **Own primary source** — direct interview, FOIA/access request,
datasets Vroeg has published.
6. **Verified secondary source** — Reuters, AP, AFP, Bloomberg, FT,
NRC, De Tijd, L'Echo, Politico Europe, EURACTIV — with direct link
to the original publication.
7. **Unverified secondary source** — blogs, press releases, conference
talks without transcripts. Permitted **only** as support for
secondary claims (background, colour), never as the sole source for
material assertions.
Every claim has at least one source from level 1-5; material numbers
have at least two independent sources from level 1-4.
Falsifier protocol
Every Vroeg article carrying a testable prediction also publishes:
- the **claim** — one sentence, unambiguous
- the **deadline** — a specific date on which the claim becomes testable
- the **falsification criterion** — what evidence would constitute proof
that the claim was wrong
Predictions are recorded at `/en/voorspellingen` with status `open`.
On the deadline, the claim is tested against primary sources and moved
to `verified`, `falsified`, or `expired`.
The track-record is public and never updated retrospectively. Wrong
predictions are not deleted — that's the whole point.
AI pipeline + human validation
Vroeg articles are written by an automated editorial pipeline (source
code open at
github.com/chutapp/vroeg).
The pipeline does:
1. **Signal detection** — three parallel layers scan primary sources weekly.
2. **Multi-perspective synthesis** — 9 journalistic "modes" run in
parallel with explicit no-fabrication prompts.
3. **Citation-grounded RAG** — every claim is generated with literal
`[F#]` / `[E#]` citations to a fixed corpus.
4. **Claim auditor + self-revision** — a separate LLM pass screens
every material claim and sends the piece back for revision if
support is missing.
5. **URL HEAD verifier + entity-status check** — broken links and
retracted entities are caught before publication.
**Human validation is not optional.** From Phase 2 (scheduled Q3 2026)
onwards, the publication pipeline blocks articles not approved by at
least one volunteer reader from the reader corps.
In Phase 1 — before the reader corps is operational — every edition
is manually approved by the responsible editor (see
colophon).
Conflicts of interest
- Vroeg accepts no advertising, sponsorship, or advertorials.
- Vroeg accepts no compensation for writing about specific entities
or policy dossiers.
- Donations (Phase 5) are publicly recorded; donors have no influence
on editorial choices.
- The responsible editor publishes their current investments and board
mandates on the colophon page; for every material investment that
comes up in an article, an explicit disclosure line is added.
Correction + retraction
See
the public corrections log for the
full process. In brief:
- **Typographic correction** — silently patched.
- **Material factual correction** — publicly logged, with date,
description and new SHA-256 in the Phrack footer.
- **Retraction** — for serious error (wrong primary interpretation,
non-existent source, non-existent entity) the article is marked
with a prominent "RETRACTED" banner and stays online for archiving;
the URL returns HTTP 200 with `noindex`.
Source protection
For human sources (Phase 2+):
- Vroeg never uses identifying details (function + company + gender)
when a source asks for anonymity; we abstract to "an EU-AI-Office
insider" or "a Belgian compliance officer".
- Email sources are kept on an end-to-end encrypted channel (Signal
or ProtonMail).
- For highly sensitive tips: Vroeg receives via
SecureDrop — set up in Phase 2.
Diversity
- Vroeg aims for gender and geographic diversity in the sources cited
and in the reader corps that validates articles.
- Not just male C-suite quotes. Not just Brussels perspective.
- Coverage of the Netherlands, Wallonia, Flanders, and the three
smallest EU states (Luxembourg, Malta, Cyprus) stays actively in
scope despite their limited demand.
## Independence
Vroeg is not an organ of EU institutions, the Belgian federal
government, or any private AI company. Vroeg receives no institutional
funding with editorial strings.