Closed beta · access by approval
Köprü takes a research problem and puts it in front of two hundred scientific disciplines in a form that hides what the problem actually is. Each one answers in its own structural language. What comes back is put through automatic checks, and only what survives is turned into a pre-registered, executable experiment and run.
At a glance
01 — The problem
A problem that has resisted a field for a decade has been asked, over and over, in the same terms. Ask a specialist and you get the field’s memory back: the familiar framings, the familiar dead ends, the familiar citations. Ask a language model and you get the same memory, faster and more fluently written.
What is missing is not effort or literature. It is a genuinely different description of the same structure — the observation that this problem in cell biology has the shape of a problem control engineering solved in 1962, made by someone who was never told the problem was about cells.
That transfer happens by accident in the history of science, over decades, at conferences and in corridors. Köprü is built to make it happen on purpose, and to be honest about how often it does not.
02 — What is different
Every choice below makes the system return less. That is the trade being made deliberately: a research instrument is worth something only if what it hands back can be trusted without being re-derived.
Blindness
Each discipline receives the problem with its subject matter removed, so it cannot retrieve what its field already says about the topic. What it can do is describe the structure it is shown in its own terms. The result is a translation rather than a recollection — and a translation is testable.
Deterministic gates
A proposal that connects two fields only because they share a word, that leans on a citation which does not exist, or that designs an experiment whose conclusion is already assumed in its setup, is rejected. Those rulings do not depend on a model being in a good mood, and confident prose does not move them.
Honesty
When the evidence does not support a conclusion, the run ends without one. A system obliged to always produce an answer will always produce an answer, and that is precisely what makes it useless for research. Köprü is allowed to come back empty-handed, and regularly does.
Traceability
What went in, what each stage decided, what was discarded and on what grounds — recorded, and fixed once written. You can hand a colleague the identifier and they can retrace the same path. A finding you cannot retrace is not a finding yet.
03 — Breadth
Most of those two hundred will produce nothing usable for any given problem, and the gates remove most of what is left. That is expected, and it is the reason for the breadth: the value is in the handful that recognise a structure the original field has no word for, and there is no way to know in advance which handful that will be.
Breadth without filtering would just be two hundred ways to be wrong at once. Breadth is only useful because the filtering is strict, and the filtering is only affordable because it is automatic.
asked, blind
carried past the gates
Illustration, not a measurement
04 — Restraint
The failure mode of every research assistant built on language models is the same: asked a question at the edge of the evidence, it produces something that reads exactly like an answer. Nothing in the output tells you which kind you received.
Köprü closes that gap by making the empty result a first-class outcome. A run can finish, having cost real time and real compute, and report that nothing survived. That is not a failed run. It is the instrument telling you the truth about what it found.
05 — Access
We are opening the studio to a small number of researchers at a time, because every early run gets read by a person. If you have a problem that has been stuck for years and you can say precisely what would count as progress on it, that is exactly the kind of question we want in the queue.
Sign in with Google to put your name in. There is no form and no pricing page yet.
What happens next
Access is granted by approval, so please do not read the wait as a fault in the site.