EIT AI Entrepreneurs Lab 2026
An autonomous AI newsroom for Spain: it groups every outlet's coverage of the same event, checks each figure against the official body that produces it, and publishes a synthesis in which every sentence links to its evidence. When the evidence isn't there, it says so.
Sergio López Rapado (CTO) & Marta Fernández López (product) · invariantchronicle.com
The problem
Ten outlets reprinting the same wire story look like ten independent confirmations. They are one.
Headlines built on different metrics, periods or territories look contradictory when they aren't — or agree when they shouldn't.
Finding the INE series, the Eurostat table or the article of the Official Gazette behind a claim takes time and expertise most readers don't have.
Generic summarisers average narratives together and can manufacture an agreement that no evidence supports. Human fact-checkers can't cover the volume.
For whom: readers who want to know what is actually supported; and, as a B2B layer, newsrooms, fact-checkers, researchers and communication teams.
The solution
Two editions a day. No human approves each piece; humans own the rules, the sources and the rights policy.
Why AI is central
Thousands of articles a day, each with dozens of claims, matched to the right statistical series: no human team can do that for a whole country.
LangGraph agents search official catalogues (INE, Eurostat, Banco de España, BOE, EUR-Lex, OECD, World Bank…) and read the documents they find. Gemini extracts, reasons and writes.
No verdict without retrieved evidence. No invented sources: every ID must come from a tool. External text can't change the rules (prompt-injection tests). Abstaining is a correct outcome.
Gold sets per task, paired shadow runs before any model change, a cost cap per day, and every model call recorded with its tokens and price.
What we have built
One full pilot day, 1 October 2026; manual review 1–2 October 2026. Private pilot: built and run by the two of us, part-time, with no funding.
What we have learned
21 of 2,908 pilot stories. Most key facts are events (51%) and statements (13%), for which we have no investigator yet.
Reading the source a claim cites works best (≈ 3% useful verdicts); matching statistical series barely works (0.3%) and costs the most. So we measure by input, not by outcome.
The “same fact” judge reaches 0.58 precision against a 0.95 bar. We test better models in paired runs before switching.
An RSS feed is not a licence. Public launch waits for licensed press content; official data (INE, Eurostat) is reusable today.
Measured 21 September – 2 October 2026 in our own pipeline.
If it works
A free, ad-free publication where every figure opens the data point behind it.
API and data on claims, provenance and evidence for newsrooms, fact-checkers, researchers and communication teams — and alerts on how a figure travels across the press.
Country packs are data, not code: a second EU country with strong official APIs, then cross-country and multilingual stories.
What we want from the Lab
The team
Cloud solutions architect in data and generative AI, 10+ years (Inditex, MAPFRE with Google Professional Services, EY). Designed and built the whole platform and runs it. M.Eng. Telecommunications Engineering.
Dental surgeon, Master's graduate. Tests every version from the reader's side so that anyone can tell what is known and what isn't. Volunteer with the Red Cross and Cáritas.
We have shipped before: MaSer ONDA, our free app against unwanted loneliness, live on Google Play since September 2026. Contact: [email protected]