Pieza independiente
The laboratory that watches itself
Movement 5
1847–2026
I
The Vienna General Hospital had two maternity wards and the women knew which was which.
The first was staffed by doctors and medical students. The second by midwives. The first killed between ten and eighteen per cent of the women who gave birth in it. The second, between two and four. The women knew: they begged to be taken to the second, and some preferred to give birth in the street rather than go into the first. They called it street birth, and the records show that the women who gave birth on the pavement survived more often.
Ignaz Semmelweis, twenty-eight years old and an assistant on that ward, spent two years looking for the difference. He tried everything. He changed the women's posture, he changed the ventilation, he had the priest's bell rung down a different corridor so the noise would not frighten them.
In March 1847 a friend of his died. Jakob Kolletschka, a professor of forensic medicine, cut himself with a scalpel during an autopsy and died with the same signs as the women on the first ward. There was the difference. The doctors came from the dissecting room and went straight to the births, without washing their hands [still wearing the apron, as well]. The midwives did no autopsies.
Semmelweis imposed hand washing with calcium chloride. Mortality on the first ward collapsed within months.
They got rid of him. His post was not renewed, he went back to Hungary, wrote a furious book calling the obstetricians of Europe murderers, and in 1865 he was committed to an asylum. He died there two weeks later, at forty-seven, of an infection in a wound the guards gave him.
II
It usually gets told as the story of a genius rejected out of envy, and that is the half that is false.
Semmelweis had the right data and the wrong mechanism. He talked about cadaverous particles, which do not exist. Germ theory would arrive years later, with other names and other laboratories. His explanation was unacceptable because it was incorrect, even though his measure was right and his figures unanswerable.
There was something worse, and his colleagues never said it out loud. Accepting it meant accepting that they had spent years killing women with their own hands. The resistance was not only academic vanity. It was the moral cost of being right against yourself.
Hold on to both, because the rest of the movement comes out of them. The right data is not enough. A wrong mechanism can live alongside a result that works.
III
Eighty years after hand washing was accepted, the cycle Semmelweis ran by hand began to be automated in sections.
In February 2025 a system appeared that generates hypotheses, criticises them and ranks them using several agents arguing with each other. In May, another that combines program generation with an automatic evaluator and finds better algorithms than the known ones for certain problems. In December a lab announced a robotic facility for synthesising and characterising materials at industrial volumes. In April 2026 a federal agency put thirty-four million dollars into autonomous catalyst laboratories. Between March and June, systems appeared that automate cell analysis, reasoning over scientific models, and protocol design. In September, one that turns a published paper into an agent able to run its own tools.
Lined up, those milestones draw a closed circuit.
No single piece is autonomous yet. All of them together are starting to look like it.
It is worth saying how big each thing is, because the headline makes them equal and they are very different. A system that ranks hypotheses is software; a facility that synthesises and characterises hundreds of compounds a day is a building with technicians, maintenance and shifts. The first deploys in an afternoon. The second takes years and there are very few [and nearly all belong to two or three organisations].
IV
I spent an afternoon pipetting next to an automated station, to understand the difference from hands.
Ninety-six wells, four plates, a protocol anyone can learn in an hour and nobody can hold for four. The room is at twenty-one degrees and smells of ethanol and new plastic. At twenty to four I skipped a column and I knew half a second later, from the movement, before I looked.
I threw the plate away — ninety-six wells in the bin for half a second of movement.
What I did next is the part that counts. I wrote it in the notebook with the time, the plate number and the reason, because that is how a decent laboratory works and because in three months somebody might wonder why a replicate was missing. That note, written by hand in two lines, is the minimum version of everything the automated facilities have to solve at another scale. The consumable costs little and the reagent costs a fair amount. What was expensive was the hour and a half before it.
The machine next to me does not skip columns. It also notices nothing. If the reagent in reservoir three has degraded, it distributes it with admirable precision across all three hundred and eighty-four positions, and the report will come out clean, with tight error bars and a pretty curve.
There is the asymmetry I took away from that afternoon. Human error is noisy, local, and usually noticed by whoever makes it. An automated system's error is silent, coherent, and spreads with exactly the efficiency with which a correct result spreads.
V
So the thesis of this movement is not about speed.
When the hypothesis, the experiment and the reading of the result are all produced by the same system, error stops being human and becomes systematic.
A bias in an instrument used to contaminate a laboratory. Today it contaminates a corpus, and the corpus trains the model that proposes the next experiment. Semmelweis took eighteen years to convince nobody with a false mechanism. A false mechanism inside an automatic loop does not take eighteen years to spread. It takes as long as the compute takes, and it arrives with impeccable internal coherence [charts included].
The dangerous thing is not that the machine gets it wrong. It is that it gets it wrong consistently and cheaply.
The known antidote is unglamorous and it is called provenance: keeping where each datum came from, with which instrument, in which reagent lot and under which model version. Without that, a suspect result cannot be traced to its origin, and all that is left is to repeat everything. The laboratories that do it well do not publish faster; they publish less often and they retract less.
“An error repeated with precision stops looking like an error. It looks like a law.”
VI
There is an exception and it deserves naming, because it explains why some branches run faster than others.
In mathematics verification is cheap. A formal proof is checked by a program, which is why the loop spins fast there and produces new things: constructions, bounds, algorithms with a computable evaluator. In July 2024 a system reached silver-medal level at the international olympiad. In 2025, another produced constructions improving on known results.
In a wet lab the verifier is the world, and the world takes time. A cell divides when it divides. A protein folds or it does not. A catalyst has to be synthesised, and synthesising it is a week.
Out of that comes the test for this branch, which is not the concrete of the previous movement nor the interventions per shift of the other. Here the strategic asset is the verifier: whoever has cheap and honest verification will move faster, and whoever has only generation will produce plausibility at industrial scale.
VII
It is worth looking closely at what a co-scientist actually does. The name promises more than it delivers and, at the same time, less than it weighs.
It generates hypotheses from the literature, sets them against each other in something like a tournament, has other agents criticise them, and returns a ranked list. That is a lot. It is not discovering. It runs no experiments. It measures nothing. Its product is a priority queue, and the value of that queue depends entirely on two things: the quality of the input corpus and the criterion by which the output is ranked.
That makes it an allocation instrument, not a discovery one. It decides where the expensive time goes [the lab's, the people's, the reagent's], and that decision, taken a thousand times, shapes a whole field without ever having run an experiment.
Out of that comes a question nobody asks in the presentations. Which hypotheses stay systematically at the bottom of the queue? The ones contradicting the corpus consensus, the ones needing odd instrumentation, and the ones that cannot be evaluated cheaply. Historically, those are three of the places the science that changed something came from.
Semmelweis would have been at the bottom of that queue. His hypothesis contradicted the consensus of his corpus, his instrumentation was a basin and his mechanism was incorrect. The only thing in his favour was a table of deaths by ward and by month, measured over two years, which is the kind of evidence a queue ranked by theoretical plausibility does not know how to score.
VIII
Against myself, twice.
First. I have used the empty months in the ledger as proof that a branch was not accelerating, and in this one the test turns on me, because almost all the milestones are published papers. A paper is an announcement with a bibliography. A robotic facility synthesising hundreds of compounds a day is a different thing, and there are few of those and they are recent.
Second. Calling physical validation a bottleneck sounds like engineering and is a category error. 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.
IX
What remains is what could not be seen in 1847.
The calcium chloride worked while the explanation was false. Hands came clean of something Semmelweis described badly, and the women stopped dying all the same. Technique got ahead of theory, which is what it almost always does, and that is why it works even when we understand it backwards.
An automated loop will do the same, faster. It will produce things that work without anyone knowing why, and some of them will be valuable.
The difference is in who goes mad. Semmelweis was right and could not prove it in the language of his time; he shouted for eighteen years and ended in a cell.
A system that is wrong coherently does not shout. It publishes.