MODUS
Every project here begins with a theorem, and the theorem decides what the project is allowed to claim about itself.
A research lab working on web3. We take results from mathematics and philosophy and wire them into products as constraints — including the constraint that says a system may not act yet.
METHOD
Most technical claims in this industry are unfalsifiable by construction. A model reports the score that flatters it. A protocol describes the case where it works. Nothing in the design makes it possible to be wrong in public, so being wrong costs nothing.
We work from the other end. Before a project is built we find the result that bounds it, and we put that bound inside the product — not in the documentation, where it can be ignored, but in the function that decides whether the product is permitted to act.
A theorem earns its place by forbidding something. If it only supplies vocabulary, it is decoration and we drop it. The test is the same every time: name the outcome this result rules out, and point to the line of code where that prohibition lives.
Work is published while it is still going badly. A lab that shows only finished results is showing you a brochure, and the failures are the reason the successes mean anything.
PROJECTS
Émile
An autonomous research agent on Robinhood Chain. It reads every token that clears $10,000 peak market capitalisation and estimates whether that token goes on to reach $30,000.
Basis. Vapnik–Chervonenkis uniform convergence, which bounds how far a measured score can sit above the true one for a model of given capacity on a sample of given size. Vapnik and Chervonenkis, 1971.
What it forbids. Émile may act on the measured score minus the penalty, and on nothing else. The subtraction happens before the number is displayed, there is no path around it, and today the result is below zero. So he does not act.
QUEUE — UNDER CONSIDERATION, NO DATES
ARROW, 1951
No voting rule satisfies every fairness criterion at once, and on-chain governance keeps claiming to have found one.
A governance system that names the criterion it gives up.
GOODHART, 1975
A measure used as a target stops being a good measure, and every incentive design in this industry is a target.
Reward mechanisms that expire before they are gamed.
BRAESS, 1968
Adding a route to a network can make every traveller slower, and liquidity routers add routes by default.
A router permitted to decline a pool it was given.
CONDORCET, 1785
Crowds outperform individuals only above a competence floor, and prediction markets assume that floor is always met.
A market that reports when its own crowd sits below it.