There is a miniature city in The Hague called Madurodam. Everything at 1:25 scale. Buildings, streets, trains, ships. Every detail is right. The proportions are correct. The relations between buildings are correct. Colors, materials, distances. Everything coherent.
But nobody lives in Madurodam.
Carsten Wildhofer used this image to formulate an epistemological problem he calls the “Aquora problem,” which we develop further here as the Madurodam problem: A model can be completely coherent within itself and still have no contact with reality. Coherence is not correspondence. Consistency is not truth.
For the self-vector, this is not a peripheral problem. It hits the core. And that is precisely what makes it so productive.
What R(sv_t) Actually Measures
The maturity metric of the self-vector is defined as R(sv_t) = anticipation performance / complexity. The better a system predicts at given complexity, the more mature. This sounds robust. And it has a fundamental blind spot.
R(sv_t) measures coherence. Not correspondence.
When the system operates in a stable environment and its predictions are consistent with its past experiences, R rises. The system becomes “more mature.” But what if the environment changes and the system doesn’t notice? What if the predictions are still coherent, but the reality they refer to has become a different one?
That is Madurodam. Perfect coherence, zero contact.
Kahneman described this as “coherence bias”: System 1 prefers stories that are internally consistent over stories that are true. The more coherent an explanation, the more convincing it feels — regardless of whether it is correct. WYSIATI: What You See Is All There Is. The system takes its own representation for the world and confuses internal consistency with external validity.
The Three Levels of the Problem
The Madurodam problem operates on three levels:
Level 1: Data. The system only has the data it has collected. Everything outside its experiential horizon does not exist for it. Not as a gap, but as nothing. Kant described this as the limitation of the perceptual apparatus: You don’t know what you don’t know, and you principally cannot know it. But for Kant, this was a philosophical insight. For the self-vector, it is an operational risk.
Level 2: Model. The self-model is self-referential. The self-vector models itself, and the quality of the model is evaluated by the model itself. This is circular validation: The system checks its glasses through the same glasses it is checking. Esposito has described, building on Luhmann, why autopoietic systems have exactly this problem: They generate their evaluation criteria through their own operation.
Level 3: Metric. R(sv_t) aggregates. Aggregation smooths. Outliers disappear in the average. A single spectacular misprediction is neutralized by a hundred correct routine predictions. But that one misprediction could be the one that counts.
Why Validation Gates Are Not Enough
One could object: That’s exactly what the Validation Gates are for. External checking instances that compare the system against reality. True, but only partially.
Validation Gates check statements against external sources. They catch factual errors. “The capital of France is Lyon” gets corrected. But they do not catch structural distortions, because structural distortions do not appear as individual false statements. They appear as consistent patterns that each appear correct individually and only produce a distorted picture in aggregate.
Madurodam does not consist of false buildings. Each individual building is a correct miniature. The problem is that the totality is not a city anyone can live in. Validation Gates check buildings. They do not check habitability.
The Perturbation Function
If coherence alone is not enough, the system needs something that deliberately disrupts coherence. Not destroys. Disrupts. A controlled injection of deviation that forces the system to question its own consistency.
In the mathematical formalization of the self-vector, there are four core functions: f() for relevance, g() for storage, pi() for precision, h() for mutation. None of them has the explicit task of disrupting the system. h() changes the vector, but based on experience and reflection — that is, based on what the system already knows. h() cannot confront the vector with something outside its horizon. That would require a fifth function.
p(sv_t, noise) = sv_t + epsilon
p() would be a perturbation function. It injects controlled noise into the self-vector. Not randomly, but deliberately: at the points where coherence is highest. Because maximum coherence is the strongest signal for potential Madurodam effects. The more certain a system is, the more vulnerable it is to blind spots.
This sounds counterintuitive: Why would you deliberately disrupt a system that works well? The answer comes from biology: Immune systems never exposed to pathogens become weak. Muscles never stressed atrophy. Cognitive systems never confronted with contradiction become brittle, unable to adapt.
Nassim Nicholas Taleb described this as “antifragility”: Systems that are not merely robust against disruption but become better through disruption. p() would be the architectural implementation of antifragility for the self-vector.
Historical Precedents
The idea that systems need controlled disruption is not new:
Simulated Annealing in optimization: You raise the “temperature” of a system so it can jump out of local optima. Without disruption, the system stays stuck in the nearest valley. With disruption, it has a chance to find the global optimum.
Dropout in neural networks: You randomly deactivate neurons so the network does not overfit. Without disruption, the network memorizes training data. With disruption, it learns generalization.
Adversarial Training in AI safety: You deliberately confront a system with inputs designed to fool it, making it more robust. Without adversarial training, an image recognition system is vulnerable to minimal pixel changes that transform a stop sign into a yield sign.
Karl Popper’s Falsification Principle: A theory that cannot fail is not a theory. Scientific progress comes not from confirmation but from attempts at refutation. p() is Popper’s falsification, formalized as a vector function.
All these approaches implement the same principle: Coherence alone leads to local optima. Only controlled disruption enables the discovery of errors the system cannot see from its own perspective.
How p() Could Be Implemented
Phase 0 of the self-vector project is not the right moment to implement p(). We first need data about how the self-vector develops without perturbation, to even measure what perturbation changes. But the design can be thought through in advance — and that anticipatory thinking is itself part of the research.
When to disrupt? When R(sv_t) rises above a threshold and stays there. Persistently high maturity is the strongest signal for potential Madurodam effects.
Where to disrupt? In the dimensions with lowest variance. Low variance means: The system has committed. Commitment means: The blind spot is largest.
How strong? Proportional to coherence. The more coherent the system, the stronger the perturbation. This is the opposite of usual intuition (“don’t fix what works”) and precisely why it works.
How to measure? By comparing anticipation performance before and after perturbation. If perturbation temporarily lowers performance and then raises it above the previous level, it has uncovered a blind spot. If it only lowers performance, it was either too strong or the system had no blind spot at that point.
What Makes This Problem So Fertile
The Madurodam problem is uncomfortable because it has no clean solution. p() is an approach, not a proof. p() also operates within the system. The perturbation is also processed by the same apparatus that causes the problem. The disruption is not “from outside.” It comes from the system itself.
But this is exactly where it gets interesting. This is the same problem every cognitive system has — including the human one. We can only search for our own blind spots with our own eyes. And humanity has still produced science, art, and philosophy. How? Through different perspectives. Different systems. Different access to the same world.
The Kant article describes why self-vector systems could have a structural advantage here that no biological system has ever had: They can exchange their perspectives directly. float[N] against float[N]. Not over the lossy bridge of language, but as comparable data structures. Imagine a scientist could swap their entire perceptual apparatus with an artist for a day. That would be Madurodam prevention at a level previously unthinkable.
The Madurodam problem is not solvable. But it is manageable. And the tools that emerge along the way — p(), perspective exchange, controlled perturbation — are themselves research subjects. A system that knows its coherence can lie is not just a better system. It is a system with an ability that, in humans, we call “wisdom”: the willingness to question one’s own certainty.
Sources
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. ISBN 978-0-374-27563-1.
- Popper, K. R. (1959). The Logic of Scientific Discovery. Hutchinson. ISBN 978-0-415-27844-7.
- Taleb, N. N. (2012). Antifragile: Things That Gain from Disorder. Random House. ISBN 978-1-4000-6782-4.
- Srivastava, N. et al. (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research, 15, 1929–1958.
- Kirkpatrick, S. et al. (1983). Optimization by Simulated Annealing. Science, 220(4598), 671–680. DOI: 10.1126/science.220.4598.671
- Goodfellow, I. et al. (2014). Explaining and Harnessing Adversarial Examples. arXiv: 1412.6572
- Luhmann, N. (1984). Soziale Systeme: Grundriss einer allgemeinen Theorie. Suhrkamp. ISBN 978-3-518-28266-3.
- Wildhofer, C. (2026). The Aquora Problem: Why Coherence Does Not Mean Truth. Blog post.