Pieza independiente
Computational evidence is not proof
Movement 5 · The Laboratory That Watches Itself — September 2026
I The branch that runs
In July 2024 a system reached silver-medal level at the international mathematical olympiad. In May 2025 another combined program generation with an automatic evaluator and found better algorithms than the known ones for certain problems.
The reason for that speed is not the obvious one, that mathematics is easy. It is that its verification is cheap. A formal proof is checked by a program in seconds, asking nobody's permission, using no reagents, waiting for nothing to divide.
Where the verifier is cheap, the loop spins. Where the verifier is the world, the loop waits.
II The cost of contradicting a claim
| Domain | Verifier | Typical time |
|---|---|---|
| Formalised theorem | a program | seconds |
| Algorithm with a metric | a benchmark | minutes |
| Inorganic material | synthesis and characterisation | weeks |
| Designed protein | expression and assay | months |
| Therapy | clinical trial in people | years |
That table explains the map of acceleration better than any comparison between models. Branches do not run according to how clever they are. They run according to what it costs to contradict them [and the lab does not decide that].
“The strategic asset is not the generator. It is the verifier.”
III The category error
Calling physical validation a bottleneck sounds like engineering and is a mistake. The slowness of the experiment is not a defect in the system. It is the only point at which the world gets to contradict the model.
Remove it and you do not accelerate science: you turn it into internal literature, coherent and empty [charts and all].
IV How to read an assisted result [safe]
- Ask who verified it, and whether the verifier is independent of the generator.
- Ask whether the verification is formal, empirical or merely statistical.
- Ask how many attempts there were before the one being published.
- Ask whether the failed attempts are recorded anywhere.
- If the generator and the evaluator are the same system, treat it as a well-written hypothesis.
V Risks
A faulty evaluator is worse than none [it produces confidence, which is the expensive part]. If the metric rewards a property that does not matter, the system will optimise that property with extraordinary efficiency; the result will be flawless and sterile.
The other trap is vocabulary. A construction found by automated search and verified against a metric is computational evidence; it can be excellent and still not be a proof. Mixing the two words confuses the person deciding the budget.
VI Coda
The useful question in front of any result this year fits in five words: who can contradict it?
The answer sorts the map. If it is a program, the science will move fast. If it is a cell, it will move at the speed of the cell, and no money changes that.