Most forecasting is unfalsifiable by construction.

"I think rates stay high." High compared to what, measured on which date, and what exactly would prove you wrong? Nothing, usually. It is a mood with a number stapled to it, and six months later it gets quietly reinterpreted into having been right all along.

For the last few months I have been building the opposite of that. It is live at intelligence.flatnine.co/theses, and there are seven theses running on it right now.

What counts as a thesis

Four parts, all of them mandatory. If any one is missing, it does not go on the board.

  1. A forward statement with a horizon. Not "the euro is weak" but "on 31 December 2027 the USD/EUR rate is still below 1.20". A date and a number you can check.
  2. A named mechanism. The causal story written out in full, so that when the thesis fails you can tell whether the conclusion was wrong or the reasoning was. Those are different failures and only one of them is interesting.
  3. Premises bound to real data. Every premise points at a numeric series or a tracked legislative instrument. Not at a feeling, not at a quote from someone confident.
  4. A falsifier, written in advance. The exact observation that kills it, committed before the outcome is known so it cannot be moved afterwards.

Then the part that matters most to me:

The strength score is computed by a nightly job. No model writes it. A language model can draft the narrative that explains a thesis, but it never touches the number. If a model scored its own theses you would just be asking it whether it agrees with itself, and it always does.

What that score means is also narrower than it looks. It is premise support, not probability. A thesis at 0.88 is not 88% likely to happen. It means the premises that back it are, right now, holding.

Two theses that contradict each other, on purpose

The board carries a claim that the EU will keep regulating AI without ever taxing its output, and also the claim that an AI output tax reaches proposed status by the end of 2027.

Both read the same counter. One of them has to resolve false.

That is deliberate, and it is the most useful thing on the board. A comfortable thesis that drifts for three years teaches nothing. A pair where one side must die produces a scored outcome on a known date, and scored outcomes are the only thing that can eventually tell me whether any of this works.

The evidence genuinely points both ways, which is why both are there:

  • For the rules, against the tax: there are 10 tracked EU instruments regulating AI systems and exactly zero instruments at any status whose base is AI output or AI compute. When Europe does tax digital activity it taxes revenue booked by foreign platforms, an administratively settled base, rather than inventing a new measure nobody has defined. Rules are cheap to pass and politically rewarding. A levy needs a definition, a measurement regime, and unanimity on tax matters.
  • For the tax: the euro area carries 87.7% debt to GDP, the EU27 old-age dependency ratio climbs about half a point a year, and old-age spending is already 10.6% of GDP. Finance ministries need bases that are visible, concentrated and mostly foreign-incident. They have done exactly this before: when the EU-level digital services tax stalled in 2018, France, Italy and Spain each legislated their own within three years. And the AI Act already built the supervisory machinery a levy could attach to, which removes the usual "we cannot measure it" objection.

The one I find hardest to argue with

Singapore out-grows Italy per head while ageing faster than Italy.

We explain Italian stagnation with demography. Old country, shrinking workforce, nothing to be done. Singapore is a controlled experiment against that story, because it is in worse demographic shape than we are: its 65+ share is rising 0.51 points a year against Italy's 0.34, and its fertility rate is 0.97 against Italy's 1.18.

If an old, shrinking population were sufficient to explain stagnation, Singapore should be the slower of the two. It grew GDP per capita 3.75% in 2025 against Italy's 0.60%, and it has out-grown Italy every year since 2023.

So the age pyramid is not the variable. What the economy does with it is: capital per worker, migration treated as an input rather than a controversy, sector mix.

The falsifier is already twitching, which is the honest part. Italy's own per-capita growth is accelerating at 0.43 points a year against a 0.4 threshold. If the gap closes, it closes from the Italian end, and that would be good news badly disguised as a dead thesis.

France is becoming Italy, from the other direction

In 2020 Italian debt exceeded French debt by 40 points of GDP. It is 21 points now.

Italy came down from 155% to 136.7%, roughly 3.7 points a year, with the deficit back from -7.1 to -3.1. France went the other way: 109.5% in 2023, 112.6% in 2024, 115.7% in 2025, on the widest deficit among the large member states.

None of that is sentiment. It is arithmetic. A primary balance near zero with nominal growth above the effective interest rate grinds a ratio down, and the reverse builds it up. Both countries have been on their own side of that line for three consecutive prints.

One thesis is about me

I invoice in euros. The cost line that grows is priced in dollars: model inference, infrastructure, SaaS tooling. No hedge, and no ability to reprice customers mid-contract.

USD per EUR went from 1.166 in August 2026 to 1.103 in November. That is 5.4% of purchasing power gone in three months on every dollar invoice, without anyone doing anything wrong.

I put it on the board because an exposure you have not written down as a falsifiable claim is just an anxiety. The thesis is deliberately set as a level holding rather than a direction continuing, because at horizons under two years an exchange rate is close to a random walk and I would rather say that out loud than pretend to forecast it.

What runs underneath

The visible part is seven theses. The machinery under them:

  • 188,568 articles ingested and stored.
  • 13,919 claims extracted from them and attached to 599 entities: companies, institutions, countries.
  • 37 macro series pulled from FRED, Eurostat and the World Bank, plus derived series, for 1,762 observations.
  • 15 EU legislative instruments tracked, with probes against the official EU law database so a new one gets noticed rather than missed.
  • A nightly check that re-reads every premise of every thesis against the latest data and recomputes the score.

What I am not claiming

Zero theses have resolved. The board needs ten resolved before any thesis is allowed to display as high confidence, and it is at zero, so everything on it is capped at medium by rule rather than by merit. The tier does not currently distinguish a 0.99 from a 0.60.

There is also a limitation worth naming, because it is the kind of thing a dashboard hides. Strength is a snapshot of today, while several of these theses require a condition to hold for years. A claim that every monthly print stays above a line for fourteen months can look perfect this evening and be dead by March. The score measures the premises now. It does not measure endurance.

The whole point is to find out. Falsification is the only part of a forecast that was ever worth anything, and most forecasts are carefully built to avoid it.

The board is here: intelligence.flatnine.co/theses. Figures as of 18 November 2026.